Paper deep dive
Bench2Dex: Benchmarking Visuo-Tactile Bimanual Dexterous Manipulation Across Dexterous Hands
Zhenjie Yang, Yideng Zhang, Dongjie Zhang, Chenyu Jiang, Xianshuai Liu, Yufeng Li, Zuhao Ge, Xingyu Jiao, Zheng Zhang, Kaiyu He, He Wang, Yuwen Zhong, Yi Deng, Muyun Jiang, Xianliang Huang, Haisheng Su, Donghang Zhang, Jian Zhang, Xue Yang, Hongyang Li, Zuxuan Wu, Yu-Gang Jiang, Xiaosong Jia, Junchi Yan
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Tactile sensing provides contact information that can be difficult to infer from vision alone, but tactile hardware for dexterous hands has not converged to a common design. Dexterous hands differ in finger structure, contact surfaces, and sensor layouts, while simulated tactile signals still differ from measurements produced by physical sensors. These factors make it difficult to study visuo-tactile manipulation across diverse dexterous hands within a consistent experimental setting. We present Bench2Dex, a simulation benchmark for visuo-tactile bimanual manipulation across 12 dexterous hands. We adapt existing robot models with a shared simulated tactile interface that converts local contact geometry into image-like tactile observations. The interface provides a consistent observation format across different hand morphologies without attempting to reproduce the output of a specific physical tactile sensor. Bench2Dex includes 26 bimanual manipulation tasks that involve tool use, articulated-object interaction, and multi-stage manipulation, together with about 1.3K human-teleoperated demonstrations. The benchmark provides synchronized visual, tactile, proprioceptive, action, and object-state observations, together with executable task metrics. For robustness, we group seven perturbation types into invariance axis, where the correct action does not change, and equivariance axis, where the correct action changes together with the perturbation. We evaluate ACT, Diffusion Policy, pi0.5, and GR00T N1.5 on Bench2Dex and report their performance and failure modes. Bench2Dex is meant as a platform for studying visuo-tactile learning across dexterous hands. It does not assume that simulated tactile observations can replace real tactile sensing; it offers a shared setting for algorithm development while tactile hardware and simulation models are still evolving.
Tags
Links
- Source: https://arxiv.org/abs/2609.15726v1
- Canonical: https://arxiv.org/abs/2609.15726v1
Trouble viewing inline? Open PDF directly โ
Full Text
123,213 characters extracted from source content.
Expand or collapse full text
Bench2Dex: Benchmarking Visuo-Tactile Bimanual Dexterous Manipulation Across Dexterous Hands Zhenjie Yang 3* Yideng Zhang 1* Dongjie Zhang 2,5* Chenyu Jiang 2,5* Xianshuai Liu 1 Yufeng Li 1,5 Zuhao Ge 2 Xingyu Jiao 2,5 Zheng Zhang 1 Kaiyu He 1 He Wang 1 Yuwen Zhong 1 Yi Deng 1 Muyun Jiang 7 Xianliang Huang 2 Haisheng Su 1 Donghang Zhang 4 Jian Zhang 4 Xue Yang 1,6 Hongyang Li 3 Zuxuan Wu 2 Yu-Gang Jiang 2 Xiaosong Jia 2โ Junchi Yan 1โ 1 Shanghai Jiao Tong University 2 Fudan University 3 The University of Hong Kong 4 Inspire Robots 5 Zhongguancun Academy 6 COWARobot Co. Ltd 7 Nanyang Technological University * Core contribution โ Corresponding authors Contact: yangzj@hku.hk, jiaxiaosong@fudan.edu.cn [ Project Page ][ Full Code ][ Huggingface ][ ModelScope ][ Documentation ] Figure 1. Overview of Bench2Dex. Bench2Dex is a simulation benchmark for bimanual dexterous manipulation with 26 tasks, 12 robot embodiments, and about 1.3K human-teleoperated demonstration trajectories, collected across a range of objects and scenes. It supports teleoperated demonstration collection and 8 data modalities: RGB images, depth maps, joint states, object states, a shared visuo-tactile representation, 2D/3D bounding boxes, and occupancy grids. With this multimodal data and controlled domain randomization, Bench2Dex supports evaluation of policy performance, robustness, and generalization on everyday bimanual manipulation tasks. Abstract Tactile sensing provides contact information that can be difficult to infer from vision alone, but tactile 1 arXiv:2609.15726v1 [cs.RO] 14 Sep 2026 hardware for dexterous hands has not converged to a common design. Dexterous hands differ in finger structure, contact surfaces, and sensor layouts, while simulated tactile signals still differ from measurements produced by physical sensors. These factors make it difficult to study visuo-tactile manipulation across diverse dexterous hands within a consistent experimental setting. We present Bench2Dex, a simulation benchmark for visuo-tactile bimanual manipulation across 12 dexterous hands. We adapt existing robot models with a shared simulated tactile interface that converts local contact geometry into image-like tactile observations. The interface provides a consistent observation format across different hand morphologies without attempting to reproduce the output of a specific physical tactile sensor. Bench2Dex includes 26 bimanual manipulation tasks that involve tool use, articulated-object interaction, and multi-stage manipulation, together with about 1.3K human-teleoperated demonstrations. The benchmark provides synchronized visual, tactile, proprioceptive, action, and object-state observations, together with executable task metrics. For robustness, we group seven perturbation types into invariance axis, where the correct action does not change, and equivariance axis, where the correct action changes together with the perturbation. We evaluate ACT, Diffusion Policy,ํ 0.5 , and GR00T N1.5 on Bench2Dex and report their performance and failure modes. Bench2Dex is meant as a platform for studying visuo-tactile learning across dexterous hands. It does not assume that simulated tactile observations can replace real tactile sensing; it offers a shared setting for algorithm development while tactile hardware and simulation models are still evolving. All code for training, inference, and teleoperation is open-sourced. 1 Introduction Dexterous manipulation is a core capability for embodied agents: it lets robots interact with tools, articulated objects, and everyday environments through direct physical contact [1,2,3,4,5]. Driven by large-scale teleoperation datasets [6,7,8,9], simulation benchmarks [4,10,11,12,13,14,15,16], cross-embodiment data efforts [17,18], and vision-language-action policies [19,20,21,22], robot learning has made steady progress in scale, task coverage, and reproducibility. However, most manipulation benchmarks and datasets are still built around one fixed robot platform, one gripper or hand design, or a narrow set of sensing modalities [4,6,7,10,11,12,13,14]. Some large datasets now cover multiple embodiments, but their evaluation protocols are not built to compare dexterous manipulation policies across different arm-hand designs and contact surfaces [17, 18, 23]. This gap matters most for visuo-tactile learning. Vision-based tactile sensors output image-like contact signals that depend closely on fingertip shape and sensor layout [24,25,26,27,28]. As a result, tactile signals recorded on one hand do not transfer directly to another hand, and, to our knowledge, no existing benchmark supports visuo-tactile data collection across multiple dexterous hands under one setting [24, 29, 30, 31]. A benchmark that spans many embodiments is useful for more than coverage. Different hands have different kinematic limits, finger designs, and contact patterns, and these differences affect grasp strategy, tool use, and long-horizon task execution [1,2,29,30,32,33]. A policy that works on one hand and fails on another is not necessarily a weaker policy โ it may simply not have been exposed to that handโs shape during training. In the same way, a shared visuo-tactile setting helps check whether a policy makes general use of contact information, rather than fitting to one handโs tactile layout [24, 25, 26]. Large-scale robot data [6,7,17,18], dexterous grasping and bimanual manipulation [1,2,32,33,34], and teleoperation systems [8,9,29,30] have each advanced on their own, but largely as separate lines of work. To our knowledge, there is still no benchmark that combines diverse bimanual dexterous embodiments, teleoperated simulation, cross-embodiment visuo-tactile data collection, long-horizon tool-use tasks, executable progress evaluation, and generalization testing in one setting [24,29,30,31,32,33,34,35,36]. To address this gap, we introduce Bench2Dex, a simulation benchmark for teleoperated bimanual dexterous manipulation built on Isaac Lab [37,38]. Bench2Dex covers 12 robot embodiments and 26 long-horizon tasks built around tool use, multi-stage interaction, and daily manipulation scenarios. Human operators collect 1.3K teleoperated trajectories in simulation, and the benchmark records synchronized observations across eight modalities, including visual, geometric, proprioceptive, action, and visuo-tactile signals. Each task has a structured scene description and is scored with executable success and progress conditions. Together, these parts connect teleoperated data collection, multimodal observation, and evaluation within one framework. 2 Table 1. Comparison of Bench2Dex with Representative Manipulation Benchmarks. BenchmarkTask NumEmbodimentTool/Device UseDexterous HandTeleoperationVision-Based TactileArticulated LIBERO [13]1301โโโโ RLBench2 [39]131โโโ RoboCasa [4]1001โโโโโ DROID [6]861โโโ DexMimicGen [33]93โโโโ RealMirror [40] 51โโโโ RoboTwin 2.0 [31]505โโโ RoboMIND 2.0 [41]7396โโโ MuJoCo Manipulus [42]161โโ RoboCasa365 [43]3651โโโโโ BiCoord [44] 181โ DexJoCo [34]112โโโ DexVerse [36]191โโโ Bench2Dex (Ours)2612โ We group the seven robustness perturbation types into two kinds. Tabletop texture, lighting conditions, scene background, camera pose, and distractor objects change the input but not the task: the object and the goal are unchanged, so the correct action should stay the same, and a drop in success rate reflects sensitivity to nuisance factors rather than a harder task. Object pose and table height change the task geometry: the correct action should change together with the perturbation, so success here instead reflects whether the policy adapts correctly. We refer to the first group as invariance axis and the second as equivariance axis, following how these properties are defined for learned policies more generally. Reporting the two groups separately, rather than as one aggregate robustness score, lets a drop in success rate be traced to nuisance sensitivity or to genuine task generalization, instead of being folded into a single number. Dexterous-hand hardware has not converged on a common finger or sensor design, and the tactile interface in Bench2Dex does not reproduce the output of any specific physical sensor. We do not treat this as a reason to wait. A shared simulation setting lets the community study visuo-tactile perception and cross-embodiment dexterous manipulation under matched tasks and conditions now, and the interface can be revised as tactile hardware and simulation models mature. In sum, our main contributions are as follows: โขWe introduce Bench2Dex, a simulation benchmark for bimanual dexterous manipulation featuring 12 robot embodiments, 26 long-horizon tasks, and 1.3K human-teleoperated demonstrations. We open-source the teleoperation systems for all embodiments. โข We develop a unified visuo-tactile interface that maps local contact geometry to a common image-like tactile representation across 12 dexterous hands with diverse finger structures and sensor layouts. Together with vision, proprioception, and action signals, it provides synchronized data across eight modalities. โข We establish an evaluation suite covering task completion, stage-level progress, and execution quality, with robustness tests spanning seven perturbation types along two axes: invariance, where correct actions remain unchanged, and equivariance, where they transform with the perturbation. โขWe benchmark ACT, Diffusion Policy,ํ 0.5 , and GR00T N1.5 on Bench2Dex, highlighting the gap between matched scenes and controlled scene perturbations in bimanual dexterous manipulation. 2 Related Work 2.1 Manipulation Benchmarks and Datasets Simulation benchmarks such as Meta-World [10], RLBench [11], robosuite [45], CALVIN [12], LIBERO [13], and ManiSkill3 [46] provide standardized evaluation for multi-task manipulation, while MimicGen [47] and RoboTwin [48] reduce demonstration cost through data generation. Large-scale real-world datasetsโ BridgeData V2 [7], Open X-Embodiment [17], DROID [6], RoboMIND 2.0 [41], and AgiBot World [49]โhave fueled generalist policies such as Octo [21], OpenVLA [22], andํ 0 [50]. Recent robustness benchmarksโ RoboCasa [4], RoboCasa365 [43], GemBench [51], THE COLOSSEUM [52], RoboTwin 2.0 [31], RL- 3 Bench2 [39], and MuJoCo Manipulus [42]โtest controlled distribution shifts. This line of work relies on parallel-jaw grippers, so it provides no multi-finger dexterous hand or vision-based tactile support, and its evaluation protocols do not compare dexterous hand morphologies or visuo-tactile sensing configurations. 2.2 Dexterous and Bimanual Manipulation Benchmarks A separate line of benchmarks targets dexterous or bimanual manipulation specifically, but each covers only part of the properties in Table 1. Bi-DexHands [32] focuses on reinforcement learning with dual Shadow Hands but provides no visual observations, and BiCoord [44] studies bimanual coordination without dexterous hands. DexMimicGen [33] generates demonstrations for 3 dexterous embodiments across 9 tasks, but does not include tool use or tactile sensing. RealMirror [40] supports a single dexterous embodiment across 5 tool-use tasks, without tactile sensing. DexJoCo [34] supports 2 embodiments and 11 tasks with tool use, but omits tactile data. DexVerse [36] covers 19 tool-use tasks on a single dexterous embodiment with teleoperated demonstrations, but likewise does not include tactile sensing. None of these benchmarks combines multiple dexterous embodiments with tactile sensing. 2.3 Dexterous Data Collection and Teleoperation Foundational work on Adroit [53], OpenAI in-hand manipulation [54], and DexYCB [55] established dexterous hands as challenging platforms for robot learning. Recent efforts scale dexterous data through synthetic grasps [56,57], generative demonstrations [33], hand-motion reconstruction from egocentric videos [58] and human-to-robot transfer [59,60,61]. Teleoperation systems such as Mobile ALOHA [9], ALOHA 2 [62], UMI [63], and FastUMI [64] advance bimanual data collection but use parallel-jaw grippers, while AnyTeleop [29] and From One Hand to Multiple [30] address cross-hand retargeting at the algorithmic level. These methods improve how dexterous demonstrations are generated or collected, but each is evaluated on its own custom setup rather than a shared benchmark spanning multiple embodiments. 2.4 Tactile Sensing and Benchmarking for Manipulation Vision-based tactile sensors, such as GelSight [25,65], DIGIT [26], TACTO [27], and Taxim [28], convert local contact deformation into image-like tactile observations, enabling robots to reason about contact states that are difficult to infer from external vision alone. Recent studies have used tactile feedback for insertion, grasp adjustment, and contact-rich manipulation [66,67,68], while newer work further explores reusable tactile skins [69], self-supervised touch representations [70], visuo-tactile pretraining and policy learning [24,71,72], and tactile-conditioned diffusion or vision-language-action policies [73, 74, 75, 76, 77, 78]. Complementary benchmark efforts have begun to standardize tactile evaluation at different levels. EgoTactile pairs egocentric video with full-hand pressure supervision, RCT evaluates contact-sequence-aware generalization across materials and sensors, HT-Bench targets full-hand tactile representation learning over 226 tasks, and HRDexDB aligns human and robot grasp sequences for cross-embodiment study [79, 80,81,82]. At the closed-loop policy level, roto 2.0 evaluates tactile-only reinforcement learning across four dexterous morphologies, TactiDex measures physically grounded contact in real-world dexterous manipulation, and SoftVTBench introduces goal- and safety-aware evaluation for visuo-tactile deformable- object manipulation [35,83,84]. Among the benchmarks in Table 1, RoboMIND 2.0 [41] is the only other one with vision-based tactile support, but it uses parallel-jaw grippers across six embodiments, so its tactile signal is not organized around a shared representation for heterogeneous dexterous hand morphologies. These efforts address complementary slices of tactile learning, but do not jointly target a common tactile representation spanning heterogeneous bimanual dexterous embodiments and long-horizon tasks. Bench2Dex instead focuses on a unified cross-embodiment tactile representation by reconstructing tactile contact surfaces from different hand meshes and converting contact depth into a common surface-aligned tactile-map format, building on geometry-consistent penetration-depth encoding [85]. This enables consistent tactile data acquisition and evaluation across diverse dexterous embodiments. 4 As summarized in Table 1, no existing benchmark jointly supports diverse bimanual dexterous embodi- ments, teleoperated demonstration collection, vision-based tactile sensing, tool use, and articulated-object interaction. Bench2Dex fills this gap with a single simulation pipeline that unifies multiple bimanual dexterous embodiments under one teleoperation interface, records synchronized multimodal observations including tactile data, and evaluates policies under controlled distribution shifts. 3 Bench2Dex Benchmark Bench2Dex is designed as a full benchmark pipeline rather than a task collection alone. As shown in Figure 1, it connects task construction, human teleoperation, multimodal data collection and executable evaluation into a unified simulation framework. This section describes the benchmark components used to generate data and evaluate policies. 3.1 Online Teleoperation Setup Bench2Dex uses human teleoperation to collect demonstrations for long-horizon bimanual dexterous manipulation. The teleoperation system is integrated directly into the Isaac Lab simulation loop, so each episode is recorded together with the commanded action, robot state, object state, camera observations, task metadata, and success or metric signals. The recording pipeline is designed to introduce minimal per-step overhead, preserving the responsiveness required for real-time human teleoperation. Teleoperation therefore serves not only as a data collection tool, but also as a practical feasibility check for whether a task can be executed under a given robot embodiment and scene configuration. The operator controls the robot through a Manus glove and an ARKit wrist-tracking stream. The Manus runtime provides a 25-node hand skeleton through shared memory; the system converts it to 21 MediaPipe-style hand keypoints and uses DexPilot retargeting to solve target joint angles for the active robot hand. The same retargeting interface is configured for twelve dexterous hand embodiments, giving all supported hands a unified teleoperation and data-collection protocol. Arm motion is controlled separately: ARKit provides wrist translation and orientation cues, while a Pinocchio-based closed-loop inverse-kinematics controller maps the desired wrist pose to the corresponding arm joint targets. The hand and arm targets are then merged into a single absolute joint-position command for the full bimanual robot. During recording, Bench2Dex stores the action to be executed at the next simulation step, followed by the resulting post-step observations and evaluator state. Before starting or saving a trajectory, the robot is moved to a home configuration, which reduces discontinuities between demonstrations and resets the teleoperation filters and wrist anchors. 3.2 Offline Multimodal Data Acquisition The lightweight online recording described above captures only the essential motion stream. The remaining modalitiesโmulti-view RGB-D images, object bounding boxes, occupancy labels, and surface-aligned tactile observationsโare generated by an offline replay pipeline that reconstructs each episode in simulation and renders the full sensor suite. This decoupling of teleoperation from expensive rendering is a key design choice: the human operator session is kept short, while the replay step can be parallelized, re-run with updated sensor configurations, or selectively applied to a subset of episodes. Replay pipeline.For each recorded episode, the replay pipeline loads the saved object initial states and robot joint trajectory, then replays the simulation with the same random seed to ensure deterministic reproduction. Multi-view RGB-D images are rendered from the six calibrated camera viewpoints (chest, overhead, stereo left, stereo right, left wrist, right wrist). Semantic and instance segmentation labels are extracted from the rendered outputs, occupancy grids are voxelized from the scene geometry, and object bounding boxes in both 2D image coordinates and 3D world coordinates are generated from the projected object meshes. The HDF5 writer keeps camera calibration, action metadata, frame validity, and task metadata together with these 5 observations, making replay, evaluation, and cross-embodiment comparison consistent across tasks and robot hands. Tactile observations are generated during the same replay pass using the surface-aligned ray-casting pipeline described in Section 3.3. Sharpa Leap RH56DFX Taxels Sampling Contact Geometry Depth Extraction Depth-to-int8 conversion u v 240 240 Hand Morphologies Library A Contact-surface Reconstruction Tactile Acquisition Pipeline Unified Tactile Representation Resolution 240 ร240 Raw Depth Map 0.0000.0000.0340.000 0.0000.1690.8350.050 0.0061.6922.7630.500 0.0000.0004.1000.000 BC D elastomer_surface.STL Constructed origin fingertip.obj Constructed origin origin left_little_rubber_2.STL Constructed u v 240 240 u v 240 240 Principle Figure 2. Unified surface-aligned tactile acquisition for diverse robotic hands. Contact surfaces are reconstructed for twelve robot-hand morphologies, and a shared surface-aligned ray-casting pipeline maps the heterogeneous hand geometries into a common image-like tactile representation. 3.3 Unified Visuo-Tactile Data Acquisition Contact is central to bimanual dexterous manipulation. After a robot grasps a tool, pushes an articulated part, or stabilizes an object with the other hand, the most informative interaction is often hidden from external cameras. Prior systems such as TACTO [27] and Taxim [28] established image-based tactile simulation, while TacMap [85] introduced a geometry-consistent penetration-depth representation over contact surfaces. Drawing on these paradigms, we develop a unified surface-aligned tactile acquisition layer for Bench2Dex. Our contribution is a cross-embodiment interface that combines reconstructed hand contact surfaces, a shared site registry, consistent encoding, and offline replay to expose heterogeneous robot hands through one tactile data representation. As illustrated in Figure 2, for each tactile siteํ , our acquisition layer uses a precomputed local surface map derived from the hand contact mesh. Each grid cell corresponds to a fixed surface point with an associated inward-facing normal. During replay, rays are cast from these surface points along the inward normals against task-object meshes, so the first-hit depth measures how far an object surface has penetrated past the nominal contact surface; the contact-consistency-filtered depth is then encoded as an 8-bit tactile imageT ํ ,ํก โ 0,..., 255 ํปรํ for siteํ at timeํก. Bench2Dex also retains the metric ray depth and a binary contact-validity mask, while the compact tactile stream applies a piecewise depth encoding and spatial smoothing. The complete signal semantics and quantization procedure are provided in Appendix C.1. This representation converts sparse geometric contact into an image-like signal that can be processed by visual encoders or fused with RGB-D, proprioception, and object-state observations. A key design goal is embodiment consistency. Bench2Dex uses one tactile acquisition paradigm for the twelve bimanual robot-hand embodiments included in the benchmark, while preserving the local contact 6 geometry of each hand. To achieve this, we rebuilt the tactile contact-surface meshes for each benchmark hand and converted them into surface point-and-normal assets, provided in Appendix C. A tactile registry then maps each robot embodiment to its tactile site names, surface assets, resolution, normal convention, and site-body attachment rules, avoiding hand-specific branching in the data collection code. 3.4 Datasets and Policy Dataset. Bench2Dex contains about 1.3K human-teleoperated demonstrations across 26 long-horizon tasks and 12 bimanual dexterous embodiments. Each trajectory is replayed into a unified HDF5 record with synchronized RGB-D observations, joint states and actions, object states, surface-aligned tactile maps, 2D/3D bounding boxes, and occupancy grids. Policy. We evaluate ACT [8], Diffusion Policy (DP) [86],ํ 0.5 [87], and GR00T N1.5 [88]. ACT and DP are trained from scratch using multi-view RGB and joint-state observations, whereasํ 0.5 and GR00T N1.5 are fully fine-tuned from pretrained checkpoints. For the pretrained policies, state/action projections are adapted to each embodiment where required, with shape-incompatible or newly added parameters initialized randomly. Complete training configurations are provided in Appendix F. 3.5 Evaluation Metrics Bench2Dex uses a unified evaluation suite that separates primary benchmark scores from task-specific diagnostics. Unless stated otherwise, we use a reach-and-stop protocol: each task provides an executable terminal predicate, and an episode terminates successfully only after that predicate remains true for its configured dwell time (0.5 s by default). Rollouts otherwise terminate at the evaluation budget or when evaluation cannot proceed. The protocol, step budget, dwell time, resolved seed, and sampled generalization parameters are retained for each episode. The primary completion metric is stable success rate, SR = 1 ํ ํ โ๏ธ ํ=1 ํ ํ ,(1) whereํ ํ = 1only when the terminal predicate stably holds, rather than at a transient frame. To characterize partial progress on long-horizon tasks, the benchmark also provides the latched stage completion rate (LSCR), LSCR ํ = 1 ํพ ํ ํพ ํ โ๏ธ ํ=1 ํฟ ํ,ํ ,(2) whereํพ ํ is the number of stages andํฟ ํ,ํ = 1if stageํis reached at any time while its declared dependencies are satisfied by previously latched or concurrently reached stages. Latching prevents progress from being erased when an early predicate is intentionally reversed by a later action, such as closing door after opening it. For efficiency analysis, the benchmark provides mean time to stable success over successful episodes, together with SR so that this conditional quantity is not interpreted independently of completion. Supported safety diagnostics include safe success rate, hard-violation rate, drop rate, tracked-object high-speed violation rate, and the fraction of executed steps containing a task-level violation. A hard violation is a configured drop or high-speed event; high speed is a severe-motion proxy rather than a contact-force or collision measurement, while joint-limit observations remain auxiliary diagnostics. For generalization analysis, episodes are grouped into theNone,Equi.,Inv., andFullchannels, for which the benchmark can compute SR, sample counts, and, when the baseline SR is nonzero, the success-rate ratio relative toNone. Additional completion, progress, efficiency, safety, grasp, tool-use, motion, and reproducibility diagnostics are described in Appendix E. 3.6 Generalization Strategy Bench2Dex evaluates whether a policy learns task-level manipulation concepts rather than memorizing a fixed simulator instance. Each evaluation keeps the semantic goal, object set, and success conditions unchanged 7 while controlling two groups of scene factors. Scene background, tabletop texture, lighting conditions, distractor objects, and camera pose are treated as Invariance factors because they alter task-irrelevant visual conditions without changing the intended task behavior. Object pose and table height are treated as Equivariance factors because they alter task-relevant geometry and therefore require corresponding changes in reaching, grasping, and contact trajectories. The protocol defines four evaluation channels. No generalization (None) exactly restores a matched anchor episode and serves as the baseline. Equivariance-only generalization (Equi.) resamples only equivariance factors while preserving the anchorโs invariant context, testing geometric adaptation. Invariance-only generalization (Inv.) resamples only invariance factors while preserving the anchorโs task geometry, testing robustness to nuisance variation. Combined generalization (Full) independently resamples both groups without an anchor, testing whether a policy can simultaneously ignore contextual shifts and adapt to new task geometry. For each episodeํ, Bench2Dex samples a generalization configurationg ํ and instantiates the executable scene as S ํ = Build(T, g ํ ),(3) whereTdenotes the task specification andg ํ stores the resolved parameters for scene background, tabletop texture, lighting conditions, distractor objects, object pose, table height, and camera pose. The resolved configuration is saved with the episode metadata, making every channel replayable and auditable. Appendix G specifies the sampling ranges and channel composition. 4 Experiments 4.1 Experimental Setup We evaluate ACT, an image-conditioned 1D U-Net Diffusion Policy variant (DP),ํ 0.5 , and GR00T N1.5 on all 26 Bench2Dex taskโembodiment settings spanning 12 embodiments (Appendix A). Each policy is evaluated with 50 rollouts under four generalization channels, yielding26ร 4ร 4ร 50 = 20,800evaluation episodes. Evaluation uses deterministic task-partitioned episode seeds and a task-specific horizon set to1.5ร the recorded expert horizon. Channel semantics follow Section 3.6:None,Equi., andInv.are aligned by episode index to matched anchors, whereasFullis sampled independently without an anchor. Cross-channel comparisons below therefore use marginal success counts and do not assume a per-instance difficulty ordering. The primary outcome is reach-and-stop stable success (Section 3.5), and each table entry is a success count out of 50 rollouts. For a fixed policy and channel, we report the equal-weight task-macro success rate over the 26 settings. Since every setting contributes 50 episodes, this rate is numerically identical to the pooled success rate over 1,300 episodes. The SR column is the equal-weight mean of the four channel-specific success rates and introduces no additional evaluation samples. 4.2 Main Results Under the matchedNonecondition, GR00T achieves 631 successes among 1,300 rollouts (48.5%), followed by ACT with 383 (29.5%),ํ 0.5 with 355 (27.3%), and DP with 168 (12.9%). GR00T is the strict observed task-level leader on 23 of the 26 settings, whileํ 0.5 leads on two and ACT ties GR00T on the remaining setting. Its matched-condition advantage is therefore broad across the evaluated tasks. The ordering becomes less concentrated under the combinedFullshift. GR00T andํ 0.5 are nearly tied in aggregate success, with 258 (19.8%) and 256 (19.7%) successes, respectively, followed by ACT with 169 (13.0%) and DP with 50 (3.8%). Despite GR00Tโs two-success (0.2 percentage-point) aggregate advantage, ํ 0.5 has the largest number of strict observed task-level leads: 12, compared with eight for GR00T and two for ACT, with four ties. Aggregate success and the distribution of task-level leaders therefore provide complementary views of performance under the combined shift. Averaged equally over the four channels, GR00T attains 28.3% success, followed by ํ 0.5 at 20.3%, ACT at 18.0%, and DP at 7.1%. 8 Table 2. Stable-success counts across generalization channels on all 26 Bench2Dex taskโembodiment settings spanning 12 embodiments. Each task entry is the number of reach-and-stop successes among 50 rollouts. None, Equi., Inv., and Full correspond to thenone,equionly,invonly, andinvequiscene profiles, respectively. SR and LSCR are the four-channel mean stable success rate and latched stage completion rate, respectively, for each taskโpolicy pair. The bottom row reports the equal-weight task-macro SR (%) and the all-task mean LSCR. TaskEmbodiment ACTDPํ 0.5 GR00T N1.5 NoneEqui.Inv.FullSRLSCRNone Equi. Inv.FullSRLSCRNoneEqui.Inv.FullSRLSCRNoneEqui.Inv.FullSRLSCR Baking Tray Prep IIWA7+Sharpa 92106.0%11.6%22002.0%7.9%71347.5%11.0%1271010.0% 29.3% Canned Food Tray Arrangement172009.5%34.6%00000.0%14.8%746712.0% 33.9%1753012.5% 37.3% Jigsaw Puzzle Assembly00000.0%12.5% 00000.0%6.3%00000.0%7.4%100005.0% 30.3% Gaming Desk SetupJAKA ZU7+DexHand0212510171031.0% 46.0%76419.0%30.7%1056613.5% 32.3%4117342759.5% 72.2% Medicine Shoebox Pack Panda+Orca 51547.5%21.5%00000.0%9.3%30334.5%21.8%1833213.0% 26.2% Tool Box Loading17108520.0% 37.5%1467616.5% 39.8%2918202345.0% 60.6%2715211740.0% 51.0% Sports Ball Cup SortPanda+Allegro40002.0%2.5%10000.5%4.4%43246.5%20.2%150219.0% 21.5% Faucet Cup Water Fill RM65+Revo2 65227.5%35.4%32124.0%14.0%51357.0%23.8%2411181333.0% 55.0% Stationery Category Sorting30001.5% 23.1%30102.0%18.5%11112.0%8.1%51003.0%16.9% Bimanual Piano Melody xArm7+Ability 60024.0%1.0%10000.5%0.5%61156.5%6.5%3082623.0% 23.5% Wine Glass Plate Balance1511131326.0% 72.0%32113.5%39.8%1179918.0% 59.2%199161328.5% 72.5% Cleaner Box Loading xArm7+LEAP 34207533.0% 76.2%113229.0%32.8%3013292448.0% 73.8%3518241747.0% 82.3% Shoebox Accessory Pack1918111230.0% 56.3%65539.5%35.2%2313232140.0% 55.5%3215181640.5% 60.3% Toilet Lid Cleaner Pour1619201334.0% 35.5%75238.5%23.5%2619151839.0% 47.0%3212252145.0% 63.3% Breadbasket Fast-Food Loading UR5+RH5DG2 1163110.5% 15.8%51214.5%7.3%1445614.5% 15.6%152109.0%13.5% Citrus Plate Loading4320142651.5% 69.9%2379622.5% 44.4%62548.5%18.9%4822242459.0% 73.1% Fridge Wine Interhand Pour844611.0% 26.4%40102.5%41.8%969615.0% 54.3%14614821.0% 58.6% Fruit Bowl Loading UR5+RH56DFX 3415131237.0% 62.0%1569618.0% 47.8%329262043.5% 71.4%41811532.5% 68.1% Screwdriver Box & Hammer165181024.5% 56.2%1859920.5% 49.0%2312171232.0% 53.5%3518222148.0% 73.7% Trash Disposal1569718.5% 30.3%1084614.0% 21.5%3114212746.5% 63.7%2610191434.5% 47.8% Fridge Fruit Shelf Sorting UR5+Shadow 10000.5%2.6%00000.0%15.1%00000.0%21.8%20001.0%19.6% Soup Serving20001.0%11.5% 00000.0%3.1%00000.0%10.0%51003.0% 13.9% Frypan Stand & Pour UR5+Schunk 2083719.0% 51.2%1694014.5% 30.0%3119161440.0% 60.0%3922301854.5% 72.3% Microwave Bowl Loading3829292962.5% 37.4% 17156421.0% 48.6%2714272144.5% 61.1%4527363370.5% 82.0% Ball Box Loading UR5+Wuji 76228.5%10.4%00000.0%7.1%8314515.0% 19.5%31311123.0% 37.1% Condiment Box Loading1233310.5% 31.8%20001.0%4.2%122121118.5% 27.7%1343110.5% 27.7% Task mean (%)29.5% 15.4% 14.0% 13.0% 18.0% 33.5%12.9% 6.3% 5.2% 3.8% 7.1%23.0%27.3% 13.2% 21.0% 19.7% 20.3% 36.1%48.5% 18.8% 26.0% 19.8% 28.3% 47.3% The LSCR results sharpen this comparison. GR00T has the highest all-task mean LSCR at 47.3%, followed byํ 0.5 at 36.1%, ACT at 33.5%, and DP at 23.0%, preserving their ordering by mean SR. Relative to their mean SRs, these LSCR values are higher by 19.0, 15.8, 15.5, and 15.9 percentage points, respectively. These gaps show that LSCR records dependency-valid stage-level progress not represented by binary stable success. This distinction is visible near the SR floor: on Jigsaw Puzzle Assembly, DP andํ 0.5 both have 0.0% mean SR but nonzero LSCR (6.3% and 7.4%), while GR00T reaches 30.3% LSCR with 5.0% mean SR. LSCR therefore distinguishes observed dependency-valid stage progress among policyโtask cases that binary SR places at or near the floor. 4.3 Generalization Patterns In the observed policy aggregates,Noneyields the highest SR for all four policies: 29.5% for ACT, 12.9% for DP, 27.3% forํ 0.5 , and 48.5% for GR00T. UnderFull, the corresponding rates are 13.0%, 3.8%, 19.7%, and 19.8%. These are descriptive marginal differences between the sampled None and Full distributions. These values summarize marginal performance under the sampled channel distributions rather than a guaranteed difficulty ordering.Nonerestores a matched scene, each isolated channel varies one factor group while retaining the other from its anchor, andFullindependently varies both groups. Because perturbation realizations vary in magnitude and composition, aFullepisode is not paired with an episode from either isolated channel, nor can it be assumed to be harder on a per-instance basis. Channel differences therefore characterize empirical sensitivity to the sampled distributions; they do not establish per-instance monotonicity, additive penalties, or interactions between the two factor groups. The realized four-channel ordering is policy dependent. For ACT and DP, aggregate SR decreases from NonethroughEqui.andInv.toFull. Forํ 0.5 and GR00T,Noneremains highest andInv.is second, butFullexceedsEqui.. Thus, combined randomization does not forceFullto be the lowest empirical aggregate, and relative sensitivity to the isolated shifts is policy dependent in the realized evaluation. UsingNoneas the matched-scene reference, the observed marginal Full-to-None success-rate ratios are 44.1% for ACT, 29.8% for DP, 72.1% forํ 0.5 , and 40.9% for GR00T. GR00T has the highest absolute success under bothNoneandFull, whereasํ 0.5 has the largest ratio. Although a lowerFullcount is not enforced 9 by the sampling design, it occurs in 89 of the 104 taskโpolicy comparisons; the remaining 15 are tied in this realized sample. Broken down by policy,Fullis lower on 25 and equal on one ACT setting, lower on 20 and equal on six DP settings, lower on 18 and equal on eight ํ 0.5 settings, and lower on all 26 GR00T settings. The isolated axes reveal policy-dependent sensitivity. ACT and DP have higher aggregate success under Equi.than underInv.(15.4% versus 14.0%, and 6.3% versus 5.2%), whereasํ 0.5 and GR00T show the reverse ordering (13.2% versus 21.0%, and 18.8% versus 26.0%). The numbers of settings on which Inv.is higher than, equal to, or lower thanEqui.are6/9/11for ACT,7/10/9for DP,18/5/3forํ 0.5 , and 16/3/7for GR00T. Separating the two axes therefore exposes differences that would be obscured by a single aggregate robustness score. 4.4 Task-Level Heterogeneity Full-condition outcomes separate the 26 settings into 12 on which all four policies record nonzero success, three on which every policy records zero success, and 11 with mixed policy outcomes. Microwave Bowl Loading has the largest across-policyFulltotal, with(29, 4, 21, 33)successes for ACT, DP,ํ 0.5 , and GR00T, respectively. In contrast, Jigsaw Puzzle Assembly, Fridge Fruit Shelf Sorting, and Soup Serving yield zero Fullsuccesses for every policy. The mixed group exposes policy-specific strengths: ACT leads Citrus Plate Loading with 26 successes,ํ 0.5 leads Tool Box Loading with 23, and GR00T leads Gaming Desk Setup with 27. On the two non-jigsaw IIWA7+Sharpa settings, onlyํ 0.5 records nonzeroFull-condition success. These contrasts show that aggregate robustness reflects heterogeneous taskโpolicy outcomes, not a uniform ordering. These results characterize stable success for the evaluated settings. The 50 rollouts per cell are evaluation episodes rather than independent retraining replicates, so the table does not estimate between-training variability. Since tasks and embodiments are not factorially crossed, task-level contrasts do not isolate embodiment effects. Stable SR alone does not distinguish partial progress, safety, or execution time. 5 Conclusion We introduced Bench2Dex, a simulation benchmark for bimanual dexterous manipulation spanning 12 robot embodiments, 26 long-horizon tasks, and approximately 1.3K human-teleoperated demonstrations. Bench2Dex combines teleoperation, synchronized multimodal data acquisition, a shared contact-geometry-based tactile interface, and executable evaluation of task completion and stage-level progress. Evaluations of ACT, Diffusion Policy,ํ 0.5 , and GR00T N1.5 using RGB and proprioceptive observations show lower aggregate success under combined scene perturbations than under matched scenes, with outcomes varying across tasks and policies. These results provide reference points for the evaluated training and execution configurations. By bringing data collection and evaluation into a common framework, Bench2Dex provides reusable components for studying bimanual dexterous manipulation and extending evaluation to tactile-conditioned policies and transfer across dexterous hands. Limitation and Outlook. Dexterous-hand and tactile-sensor designs continue to evolve, and the hand and tactile models in Bench2Dex have not been calibrated against matching physical hardware. Bench2Dex focuses on a shared simulation interface for studying visuo-tactile manipulation across diverse dexterous hands, rather than reproducing a specific physical handโsensor system. It provides a common setting for algorithm development under explicit modeling and sensing assumptions. As hardware and simulation models advance, future extensions could incorporate refined hand and tactile models, hardware-specific calibration, and physical validation to assess which findings carry over to real-world manipulation. 10 References [1]Ruicheng Wang, Jialiang Zhang, Jiayi Chen, Yinzhen Xu, Puhao Li, Tengyu Liu, and He Wang. Dexgraspnet: A large-scale robotic dexterous grasp dataset for general objects based on simulation, 2023. URL https://arxiv.org/abs/2210.02697. [2] Yinzhen Xu, Weikang Wan, Jialiang Zhang, Haoran Liu, Zikang Shan, Hao Shen, Ruicheng Wang, Haoran Geng, Yijia Weng, Jiayi Chen, Tengyu Liu, Li Yi, and He Wang. Unidexgrasp: Universal robotic dexterous grasping via learning diverse proposal generation and goal-conditioned policy, 2023. URL https://arxiv.org/abs/2303.00938. [3]Chen Bao, Helin Xu, Yuzhe Qin, and Xiaolong Wang. Dexart: Benchmarking generalizable dexterous manipulation with articulated objects, 2023. URL https://arxiv.org/abs/2305.05706. [4]Soroush Nasiriany, Abhiram Maddukuri, Lance Zhang, Adeet Parikh, Aaron Lo, Abhishek Joshi, Ajay Mandlekar, and Yuke Zhu. Robocasa: Large-scale simulation of everyday tasks for generalist robots, 2024. URL https://arxiv.org/abs/2406.02523. [5]Chengshu Li, Ruohan Zhang, Josiah Wong, Cem Gokmen, Sanjana Srivastava, Roberto Mart ฬ ฤฑn-Mart ฬ ฤฑn, Chen Wang, Gabrael Levine, Wensi Ai, Benjamin Martinez, Hang Yin, Michael Lingelbach, Minjune Hwang, Ayano Hiranaka, Sujay Garlanka, Arman Aydin, Sharon Lee, Jiankai Sun, Mona Anvari, Manasi Sharma, Dhruva Bansal, Samuel Hunter, Kyu-Young Kim, Alan Lou, Caleb R Matthews, Ivan Villa-Renteria, Jerry Huayang Tang, Claire Tang, Fei Xia, Yunzhu Li, Silvio Savarese, Hyowon Gweon, C. Karen Liu, Jiajun Wu, and Li Fei-Fei. Behavior-1k: A human-centered, embodied ai benchmark with 1,000 everyday activities and realistic simulation, 2024. URLhttps://arxiv.org/abs/2403.09227. [6]Alexander Khazatsky, Karl Pertsch, Suraj Nair, Ashwin Balakrishna, Sudeep Dasari, Siddharth Karamcheti, Soroush Nasiriany, Mohan Kumar Srirama, Lawrence Yunliang Chen, Kirsty Ellis, Peter David Fagan, Joey Hejna, Masha Itkina, Marion Lepert, Yecheng Jason Ma, Patrick Tree Miller, Jimmy Wu, Suneel Belkhale, Shivin Dass, Huy Ha, Arhan Jain, Abraham Lee, Youngwoon Lee, Marius Memmel, Sungjae Park, Ilija Radosavovic, Kaiyuan Wang, Albert Zhan, Kevin Black, Cheng Chi, Kyle Beltran Hatch, Shan Lin, Jingpei Lu, Jean Mercat, Abdul Rehman, Pannag R Sanketi, Archit Sharma, Cody Simpson, Quan Vuong, Homer Rich Walke, Blake Wulfe, Ted Xiao, Jonathan Heewon Yang, Arefeh Yavary, Tony Z. Zhao, Christopher Agia, Rohan Baijal, Mateo Guaman Castro, Daphne Chen, Qiuyu Chen, Trinity Chung, Jaimyn Drake, Ethan Paul Foster, Jensen Gao, Vitor Guizilini, David Antonio Herrera, Minho Heo, Kyle Hsu, Jiaheng Hu, Muhammad Zubair Irshad, Donovon Jackson, Charlotte Le, Yunshuang Li, Kevin Lin, Roy Lin, Zehan Ma, Abhiram Maddukuri, Suvir Mirchandani, Daniel Morton, Tony Nguyen, Abigail OโNeill, Rosario Scalise, Derick Seale, Victor Son, Stephen Tian, Emi Tran, Andrew E. Wang, Yilin Wu, Annie Xie, Jingyun Yang, Patrick Yin, Yunchu Zhang, Osbert Bastani, Glen Berseth, Jeannette Bohg, Ken Goldberg, Abhinav Gupta, Abhishek Gupta, Dinesh Jayaraman, Joseph J Lim, Jitendra Malik, Roberto Mart ฬ ฤฑn-Mart ฬ ฤฑn, Subramanian Ramamoorthy, Dorsa Sadigh, Shuran Song, Jiajun Wu, Michael C. Yip, Yuke Zhu, Thomas Kollar, Sergey Levine, and Chelsea Finn. Droid: A large-scale in-the-wild robot manipulation dataset, 2025. URL https://arxiv.org/abs/2403.12945. [7]Homer Walke, Kevin Black, Abraham Lee, Moo Jin Kim, Max Du, Chongyi Zheng, Tony Zhao, Philippe Hansen-Estruch, Quan Vuong, Andre He, Vivek Myers, Kuan Fang, Chelsea Finn, and Sergey Levine. Bridgedata v2: A dataset for robot learning at scale, 2024. URLhttps://arxiv.org/abs/2308. 12952. [8] Tony Z. Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manipulation with low-cost hardware, 2023. URL https://arxiv.org/abs/2304.13705. 11 [9]Zipeng Fu, Tony Z. Zhao, and Chelsea Finn. Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation, 2024. URL https://arxiv.org/abs/2401.02117. [10]Tianhe Yu, Deirdre Quillen, Zhanpeng He, Ryan Julian, Avnish Narayan, Hayden Shively, Adithya Bellathur, Karol Hausman, Chelsea Finn, and Sergey Levine. Meta-world: A benchmark and evaluation for multi-task and meta reinforcement learning, 2021. URLhttps://arxiv.org/abs/1910.10897. [11]Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J. Davison. Rlbench: The robot learning benchmark & learning environment, 2019. URL https://arxiv.org/abs/1909.12271. [12]Oier Mees, Lukas Hermann, Erick Rosete-Beas, and Wolfram Burgard. Calvin: A benchmark for language-conditioned policy learning for long-horizon robot manipulation tasks, 2022. URL https://arxiv.org/abs/2112.03227. [13]Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning, 2023. URLhttps://arxiv.org/abs/ 2306.03310. [14]Jiayuan Gu, Fanbo Xiang, Xuanlin Li, Zhan Ling, Xiqiang Liu, Tongzhou Mu, Yihe Tang, Stone Tao, Xinyue Wei, Yunchao Yao, Xiaodi Yuan, Pengwei Xie, Zhiao Huang, Rui Chen, and Hao Su. Maniskill2: A unified benchmark for generalizable manipulation skills, 2023. URLhttps: //arxiv.org/abs/2302.04659. [15]TriWorldBench: A benchmark evaluating triple-view embodied world models.https://github.com/ TriWorldBench/TriWorldBench, 2026. GitHub repository. [16] Yejin Kim, Wilbert Pumacay, Omar Rayyan, Max Argus, Winson Han, Eli VanderBilt, Jordi Salvador, Abhay Deshpande, Rose Hendrix, Snehal Jauhri, Shuo Liu, Nur Muhammad Mahi Shafiullah, Maya Guru, Arjun Guru, Ainaz Eftekhar, Karen Farley, Donovan Clay, Jiafei Duan, Piper Wolters, Alvaro Herrasti, Ying-Chun Lee, Georgia Chalvatzaki, Yuchen Cui, Ali Farhadi, Dieter Fox, and Ranjay Krishna. Molmospaces: A large-scale open ecosystem for robot navigation and manipulation, 2026. URL https://arxiv.org/abs/2602.11337. [17]Embodiment Collaboration, Abby OโNeill, Abdul Rehman, Abhinav Gupta, Abhiram Maddukuri, Abhishek Gupta, Abhishek Padalkar, Abraham Lee, Acorn Pooley, Agrim Gupta, Ajay Mandlekar, Ajinkya Jain, Albert Tung, Alex Bewley, Alex Herzog, Alex Irpan, Alexander Khazatsky, Anant Rai, Anchit Gupta, Andrew Wang, Andrey Kolobov, Anikait Singh, Animesh Garg, Aniruddha Kembhavi, Annie Xie, Anthony Brohan, Antonin Raffin, Archit Sharma, Arefeh Yavary, Arhan Jain, Ashwin Balakrishna, Ayzaan Wahid, Ben Burgess-Limerick, Beomjoon Kim, Bernhard Sch ฬ olkopf, Blake Wulfe, Brian Ichter, Cewu Lu, Charles Xu, Charlotte Le, Chelsea Finn, Chen Wang, Chenfeng Xu, Cheng Chi, Chenguang Huang, Christine Chan, Christopher Agia, Chuer Pan, Chuyuan Fu, Coline Devin, Danfei Xu, Daniel Morton, Danny Driess, Daphne Chen, Deepak Pathak, Dhruv Shah, Dieter B ฬuchler, Dinesh Jayaraman, Dmitry Kalashnikov, Dorsa Sadigh, Edward Johns, Ethan Foster, Fangchen Liu, Federico Ceola, Fei Xia, Feiyu Zhao, Felipe Vieira Frujeri, Freek Stulp, Gaoyue Zhou, Gaurav S. Sukhatme, Gautam Salhotra, Ge Yan, Gilbert Feng, Giulio Schiavi, Glen Berseth, Gregory Kahn, Guangwen Yang, Guanzhi Wang, Hao Su, Hao-Shu Fang, Haochen Shi, Henghui Bao, Heni Ben Amor, Henrik I Christensen, Hiroki Furuta, Homanga Bharadhwaj, Homer Walke, Hongjie Fang, Huy Ha, Igor Mordatch, Ilija Radosavovic, Isabel Leal, Jacky Liang, Jad Abou-Chakra, Jaehyung Kim, Jaimyn Drake, Jan Peters, Jan Schneider, Jasmine Hsu, Jay Vakil, Jeannette Bohg, Jeffrey Bingham, Jeffrey Wu, Jensen Gao, Jiaheng Hu, Jiajun Wu, Jialin Wu, Jiankai Sun, Jianlan Luo, Jiayuan Gu, Jie Tan, Jihoon Oh, Jimmy Wu, Jingpei Lu, Jingyun Yang, Jitendra Malik, Jo ฬ ao Silv ฬ erio, Joey Hejna, Jonathan Booher, Jonathan Tompson, Jonathan Yang, Jordi Salvador, Joseph J. Lim, Junhyek Han, Kaiyuan Wang, Kanishka Rao, Karl Pertsch, Karol Hausman, Keegan Go, Keerthana Gopalakrishnan, Ken Goldberg, Kendra Byrne, 12 Kenneth Oslund, Kento Kawaharazuka, Kevin Black, Kevin Lin, Kevin Zhang, Kiana Ehsani, Kiran Lekkala, Kirsty Ellis, Krishan Rana, Krishnan Srinivasan, Kuan Fang, Kunal Pratap Singh, Kuo-Hao Zeng, Kyle Hatch, Kyle Hsu, Laurent Itti, Lawrence Yunliang Chen, Lerrel Pinto, Li Fei-Fei, Liam Tan, Linxi โJimโ Fan, Lionel Ott, Lisa Lee, Luca Weihs, Magnum Chen, Marion Lepert, Marius Memmel, Masayoshi Tomizuka, Masha Itkina, Mateo Guaman Castro, Max Spero, Maximilian Du, Michael Ahn, Michael C. Yip, Mingtong Zhang, Mingyu Ding, Minho Heo, Mohan Kumar Srirama, Mohit Sharma, Moo Jin Kim, Muhammad Zubair Irshad, Naoaki Kanazawa, Nicklas Hansen, Nicolas Heess, Nikhil J Joshi, Niko Suenderhauf, Ning Liu, Norman Di Palo, Nur Muhammad Mahi Shafiullah, Oier Mees, Oliver Kroemer, Osbert Bastani, Pannag R Sanketi, Patrick โTreeโ Miller, Patrick Yin, Paul Wohlhart, Peng Xu, Peter David Fagan, Peter Mitrano, Pierre Sermanet, Pieter Abbeel, Priya Sundaresan, Qiuyu Chen, Quan Vuong, Rafael Rafailov, Ran Tian, Ria Doshi, Roberto Mart ฬ ฤฑn-Mart ฬ ฤฑn, Rohan Baijal, Rosario Scalise, Rose Hendrix, Roy Lin, Runjia Qian, Ruohan Zhang, Russell Mendonca, Rutav Shah, Ryan Hoque, Ryan Julian, Samuel Bustamante, Sean Kirmani, Sergey Levine, Shan Lin, Sherry Moore, Shikhar Bahl, Shivin Dass, Shubham Sonawani, Shubham Tulsiani, Shuran Song, Sichun Xu, Siddhant Haldar, Siddharth Karamcheti, Simeon Adebola, Simon Guist, Soroush Nasiriany, Stefan Schaal, Stefan Welker, Stephen Tian, Subramanian Ramamoorthy, Sudeep Dasari, Suneel Belkhale, Sungjae Park, Suraj Nair, Suvir Mirchandani, Takayuki Osa, Tanmay Gupta, Tatsuya Harada, Tatsuya Matsushima, Ted Xiao, Thomas Kollar, Tianhe Yu, Tianli Ding, Todor Davchev, Tony Z. Zhao, Travis Armstrong, Trevor Darrell, Trinity Chung, Vidhi Jain, Vikash Kumar, Vincent Vanhoucke, Vitor Guizilini, Wei Zhan, Wenxuan Zhou, Wolfram Burgard, Xi Chen, Xiangyu Chen, Xiaolong Wang, Xinghao Zhu, Xinyang Geng, Xiyuan Liu, Xu Liangwei, Xuanlin Li, Yansong Pang, Yao Lu, Yecheng Jason Ma, Yejin Kim, Yevgen Chebotar, Yifan Zhou, Yifeng Zhu, Yilin Wu, Ying Xu, Yixuan Wang, Yonatan Bisk, Yongqiang Dou, Yoonyoung Cho, Youngwoon Lee, Yuchen Cui, Yue Cao, Yueh-Hua Wu, Yujin Tang, Yuke Zhu, Yunchu Zhang, Yunfan Jiang, Yunshuang Li, Yunzhu Li, Yusuke Iwasawa, Yutaka Matsuo, Zehan Ma, Zhuo Xu, Zichen Jeff Cui, Zichen Zhang, Zipeng Fu, and Zipeng Lin. Open x-embodiment: Robotic learning datasets and rt-x models, 2025. URL https://arxiv.org/abs/2310.08864. [18]Kun Wu, Chengkai Hou, Jiaming Liu, Zhengping Che, Xiaozhu Ju, Zhuqin Yang, Meng Li, Yinuo Zhao, Zhiyuan Xu, Guang Yang, Shichao Fan, Xinhua Wang, Fei Liao, Zhen Zhao, Guangyu Li, Zhao Jin, Lecheng Wang, Jilei Mao, Ning Liu, Pei Ren, Qiang Zhang, Yaoxu Lyu, Mengzhen Liu, He Jingyang, Yulin Luo, Zeyu Gao, Chenxuan Li, Chenyang Gu, Yankai Fu, Di Wu, Xingyu Wang, Sixiang Chen, Zhenyu Wang, Pengju An, Siyuan Qian, Shanghang Zhang, and Jian Tang. Robomind: Benchmark on multi-embodiment intelligence normative data for robot manipulation. In Robotics: Science and Systems XXI, RSS2025. Robotics: Science and Systems Foundation, 2025. doi: 10.15607/rss.2025.xxi.152. URL http://dx.doi.org/10.15607/RSS.2025.XXI.152. [19]Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, Julian Ibarz, Brian Ichter, Alex Irpan, Tomas Jackson, Sally Jesmonth, Nikhil J Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Isabel Leal, Kuang-Huei Lee, Sergey Levine, Yao Lu, Utsav Malla, Deeksha Manjunath, Igor Mordatch, Ofir Nachum, Carolina Parada, Jodilyn Peralta, Emily Perez, Karl Pertsch, Jornell Quiambao, Kanishka Rao, Michael Ryoo, Grecia Salazar, Pannag Sanketi, Kevin Sayed, Jaspiar Singh, Sumedh Sontakke, Austin Stone, Clayton Tan, Huong Tran, Vincent Vanhoucke, Steve Vega, Quan Vuong, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Tianhe Yu, and Brianna Zitkovich. Rt-1: Robotics transformer for real-world control at scale, 2023. URL https://arxiv.org/abs/2212.06817. [20] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, Pete Florence, Chuyuan Fu, Montse Gonzalez Arenas, Keerthana Gopalakrishnan, Kehang Han, Karol Hausman, Alexander Herzog, Jasmine Hsu, Brian Ichter, Alex Irpan, Nikhil Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Isabel Leal, Lisa Lee, Tsang-Wei Edward Lee, Sergey Levine, Yao Lu, Henryk Michalewski, Igor Mordatch, Karl Pertsch, Kanishka Rao, Krista Reymann, Michael Ryoo, Grecia Salazar, Pannag Sanketi, Pierre Sermanet, 13 Jaspiar Singh, Anikait Singh, Radu Soricut, Huong Tran, Vincent Vanhoucke, Quan Vuong, Ayzaan Wahid, Stefan Welker, Paul Wohlhart, Jialin Wu, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Tianhe Yu, and Brianna Zitkovich. Rt-2: Vision-language-action models transfer web knowledge to robotic control, 2023. URL https://arxiv.org/abs/2307.15818. [21]Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joey Hejna, Tobias Kreiman, Charles Xu, Jianlan Luo, You Liang Tan, Lawrence Yunliang Chen, Pannag Sanketi, Quan Vuong, Ted Xiao, Dorsa Sadigh, Chelsea Finn, and Sergey Levine. Octo: An open-source generalist robot policy, 2024. URL https://arxiv.org/abs/2405.12213. [22]Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. Openvla: An open-source vision-language-action model, 2024. URL https://arxiv.org/abs/2406.09246. [23]Hao-Shu Fang, Hongjie Fang, Zhenyu Tang, Jirong Liu, Chenxi Wang, Junbo Wang, Haoyi Zhu, and Cewu Lu. Rh20t: A comprehensive robotic dataset for learning diverse skills in one-shot, 2023. URL https://arxiv.org/abs/2307.00595. [24] Binghao Huang, Yixuan Wang, Xinyi Yang, Yiyue Luo, and Yunzhu Li. 3d-vitac: Learning fine-grained manipulation with visuo-tactile sensing, 2025. URL https://arxiv.org/abs/2410.24091. [25] Siyuan Dong, Wenzhen Yuan, and Edward H. Adelson. Improved gelsight tactile sensor for measuring geometry and slip. In 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), page 137โ144. IEEE, September 2017. doi: 10.1109/iros.2017.8202149. URLhttp://dx. doi.org/10.1109/IROS.2017.8202149. [26] Mike Lambeta, Po-Wei Chou, Stephen Tian, Brian Yang, Benjamin Maloon, Victoria Rose Most, Dave Stroud, Raymond Santos, Ahmad Byagowi, Gregg Kammerer, Dinesh Jayaraman, and Roberto Calandra. Digit: A novel design for a low-cost compact high-resolution tactile sensor with application to in-hand manipulation. IEEE Robotics and Automation Letters, 5(3):3838โ3845, July 2020. ISSN 2377-3774. doi: 10.1109/lra.2020.2977257. URL http://dx.doi.org/10.1109/LRA.2020.2977257. [27]Shaoxiong Wang, Mike Lambeta, Po-Wei Chou, and Roberto Calandra. Tacto: A fast, flexible, and open-source simulator for high-resolution vision-based tactile sensors. IEEE Robotics and Automation Letters, 7(2):3930โ3937, April 2022. ISSN 2377-3774. doi: 10.1109/lra.2022.3146945. URL http://dx.doi.org/10.1109/LRA.2022.3146945. [28] Zilin Si and Wenzhen Yuan. Taxim: An example-based simulation model for gelsight tactile sensors, 2021. URL https://arxiv.org/abs/2109.04027. [29]Yuzhe Qin, Wei Yang, Binghao Huang, Karl Van Wyk, Hao Su, Xiaolong Wang, Yu-Wei Chao, and Dieter Fox. Anyteleop: A general vision-based dexterous robot arm-hand teleoperation system, 2024. URL https://arxiv.org/abs/2307.04577. [30] Yuzhe Qin, Hao Su, and Xiaolong Wang. From one hand to multiple hands: Imitation learning for dexterous manipulation from single-camera teleoperation, 2023. URLhttps://arxiv.org/abs/ 2204.12490. [31] Tianxing Chen, Zanxin Chen, Baijun Chen, Zijian Cai, Yibin Liu, Zixuan Li, Qiwei Liang, Xianliang Lin, Yiheng Ge, Zhenyu Gu, Weiliang Deng, Yubin Guo, Tian Nian, Xuanbing Xie, Qiangyu Chen, Kailun Su, Tianling Xu, Guodong Liu, Mengkang Hu, Huan ang Gao, Kaixuan Wang, Zhixuan Liang, Yusen Qin, Xiaokang Yang, Ping Luo, and Yao Mu. Robotwin 2.0: A scalable data generator and benchmark with strong domain randomization for robust bimanual robotic manipulation, 2025. URL https://arxiv.org/abs/2506.18088. 14 [32]Yuanpei Chen, Tianhao Wu, Shengjie Wang, Xidong Feng, Jiechuang Jiang, Stephen Marcus McAleer, Yiran Geng, Hao Dong, Zongqing Lu, Song-Chun Zhu, and Yaodong Yang. Towards human-level bimanual dexterous manipulation with reinforcement learning, 2022. URLhttps://arxiv.org/abs/ 2206.08686. [33] Zhenyu Jiang, Yuqi Xie, Kevin Lin, Zhenjia Xu, Weikang Wan, Ajay Mandlekar, Linxi Fan, and Yuke Zhu. Dexmimicgen: Automated data generation for bimanual dexterous manipulation via imitation learning, 2025. URL https://arxiv.org/abs/2410.24185. [34]Hanwen Wang, Weizhi Zhao, Xiangyu Wang, Siyuan Huang, He Lin, Boyuan Zheng, Rongtao Xu, Gang Wang, Yao Mu, He Wang, Lue Fan, Hongsheng Li, Zhaoxiang Zhang, and Tieniu Tan. Dexjoco: A benchmark and toolkit for task-oriented dexterous manipulation on mujoco, 2026. URL https://arxiv.org/abs/2605.16257. [35]Suting Ni, Hanbing Zhang, Zhenyu Wei, Guo Chen, Chixuan Zhang, Ye Shi, and Jingya Wang. Tactidex: A real-world tactile-guided benchmark for human-like dexterous manipulation, 2026. URL https://arxiv.org/abs/2607.09190. [36]Yunchao Yao, Zhuxiu Xu, Tianqi Zhang, Zixian Liu, Sikai Li, Zhenyu Wei, Feng Chen, Dihong Huang, Kechang Wan, Chenyang Ma, Shuqi Zhao, Shenghua Gao, Masayoshi Tomizuka, Yi Ma, and Mingyu Ding. Dexverse: A modular benchmark for multi-task, multi-embodiment dexterous manipulation, 2026. URL https://arxiv.org/abs/2607.08751. [37]NVIDIA, :, Mayank Mittal, Pascal Roth, James Tigue, Antoine Richard, Octi Zhang, Peter Du, Antonio Serrano-Mu ฬ noz, Xinjie Yao, Ren ฬ e Zurbr ฬugg, Nikita Rudin, Lukasz Wawrzyniak, Milad Rakhsha, Alain Denzler, Eric Heiden, Ales Borovicka, Ossama Ahmed, Iretiayo Akinola, Abrar Anwar, Mark T. Carlson, Ji Yuan Feng, Animesh Garg, Renato Gasoto, Lionel Gulich, Yijie Guo, M. Gussert, Alex Hansen, Mihir Kulkarni, Chenran Li, Wei Liu, Viktor Makoviychuk, Grzegorz Malczyk, Hammad Mazhar, Masoud Moghani, Adithyavairavan Murali, Michael Noseworthy, Alexander Poddubny, Nathan Ratliff, Welf Rehberg, Clemens Schwarke, Ritvik Singh, James Latham Smith, Bingjie Tang, Ruchik Thaker, Matthew Trepte, Karl Van Wyk, Fangzhou Yu, Alex Millane, Vikram Ramasamy, Remo Steiner, Sangeeta Subramanian, Clemens Volk, CY Chen, Neel Jawale, Ashwin Varghese Kuruttukulam, Michael A. Lin, Ajay Mandlekar, Karsten Patzwaldt, John Welsh, Huihua Zhao, Fatima Anes, Jean-Francois Lafleche, Nicolas Mo ฬ enne-Loccoz, Soowan Park, Rob Stepinski, Dirk Van Gelder, Chris Amevor, Jan Carius, Jumyung Chang, Anka He Chen, Pablo de Heras Ciechomski, Gilles Daviet, Mohammad Mohajerani, Julia von Muralt, Viktor Reutskyy, Michael Sauter, Simon Schirm, Eric L. Shi, Pierre Terdiman, Kenny Vilella, Tobias Widmer, Gordon Yeoman, Tiffany Chen, Sergey Grizan, Cathy Li, Lotus Li, Connor Smith, Rafael Wiltz, Kostas Alexis, Yan Chang, David Chu, Linxi โJimโ Fan, Farbod Farshidian, Ankur Handa, Spencer Huang, Marco Hutter, Yashraj Narang, Soha Pouya, Shiwei Sheng, Yuke Zhu, Miles Macklin, Adam Moravanszky, Philipp Reist, Yunrong Guo, David Hoeller, and Gavriel State. Isaac lab: A gpu-accelerated simulation framework for multi-modal robot learning, 2025. URL https://arxiv.org/abs/2511.04831. [38]Viktor Makoviychuk, Lukasz Wawrzyniak, Yunrong Guo, Michelle Lu, Kier Storey, Miles Macklin, David Hoeller, Nikita Rudin, Arthur Allshire, Ankur Handa, and Gavriel State. Isaac gym: High performance gpu-based physics simulation for robot learning, 2021. URLhttps://arxiv.org/abs/2108.10470. [39]Markus Grotz, Mohit Shridhar, Tamim Asfour, and Dieter Fox. Peract2: Benchmarking and learning for robotic bimanual manipulation tasks, 2024. URL https://arxiv.org/abs/2407.00278. [40]Cong Tai, Zhaoyu Zheng, Haixu Long, Hansheng Wu, Haodong Xiang, Zhengbin Long, Jun Xiong, Rong Shi, Shizhuang Zhang, Gang Qiu, He Wang, Ruifeng Li, Jun Huang, Bin Chang, Shuai Feng, and Tao Shen. Realmirror: A comprehensive, open-source vision-language-action platform for embodied ai, 2025. URL https://arxiv.org/abs/2509.14687. 15 [41]Chengkai Hou, Kun Wu, Jiaming Liu, Zhengping Che, Di Wu, Fei Liao, Guangrun Li, Jingyang He, Qiuxuan Feng, Zhao Jin, Chenyang Gu, Zhuoyang Liu, Nuowei Han, Xiangju Mi, Yaoxu Lv, Yankai Fu, Gaole Dai, Langzhe Gu, Tao Li, Yuheng Zhang, Yixue Zhang, Xinhua Wang, Shichao Fan, Meng Li, Zhen Zhao, Ning Liu, Zhiyuan Xu, Pei Ren, Junjie Ji, Haonan Liu, Kuan Cheng, Shanghang Zhang, and Jian Tang. Robomind 2.0: A multimodal, bimanual mobile manipulation dataset for generalizable embodied intelligence, 2026. URL https://arxiv.org/abs/2512.24653. [42]Jonathan Zamora, Daniel Seita, and Yue Wang. Mujoco manipulus: A robot learning benchmark for generalizable tool manipulation, 2025. URL https://openreview.net/forum?id=b9Ne5lHJ8Y. [43]Soroush Nasiriany, Sepehr Nasiriany, Abhiram Maddukuri, and Yuke Zhu. Robocasa365: A large- scale simulation framework for training and benchmarking generalist robots, 2026. URLhttps: //arxiv.org/abs/2603.04356. [44]Xingyu Peng, Chen Gao, Liankai Jin, Annan Li, and Si Liu. Bicoord: A bimanual manipulation benchmark towards long-horizon spatial-temporal coordination, 2026. URLhttps://arxiv.org/ abs/2604.05831. [45] Yuke Zhu, Josiah Wong, Ajay Mandlekar, Roberto Mart ฬ ฤฑn-Mart ฬ ฤฑn, Abhishek Joshi, Kevin Lin, Abhiram Maddukuri, Soroush Nasiriany, and Yifeng Zhu. robosuite: A modular simulation framework and benchmark for robot learning, 2025. URL https://arxiv.org/abs/2009.12293. [46]Stone Tao, Fanbo Xiang, Arth Shukla, Yuzhe Qin, Xander Hinrichsen, Xiaodi Yuan, Chen Bao, Xinsong Lin, Yulin Liu, Tse kai Chan, Yuan Gao, Xuanlin Li, Tongzhou Mu, Nan Xiao, Arnav Gurha, Viswesh Nagaswamy Rajesh, Yong Woo Choi, Yen-Ru Chen, Zhiao Huang, Roberto Calandra, Rui Chen, Shan Luo, and Hao Su. Maniskill3: Gpu parallelized robotics simulation and rendering for generalizable embodied ai, 2025. URL https://arxiv.org/abs/2410.00425. [47] Ajay Mandlekar, Soroush Nasiriany, Bowen Wen, Iretiayo Akinola, Yashraj Narang, Linxi Fan, Yuke Zhu, and Dieter Fox. Mimicgen: A data generation system for scalable robot learning using human demonstrations, 2023. URL https://arxiv.org/abs/2310.17596. [48]Yao Mu, Tianxing Chen, Zanxin Chen, Shijia Peng, Zhiqian Lan, Zeyu Gao, Zhixuan Liang, Qiaojun Yu, Yude Zou, Mingkun Xu, Lunkai Lin, Zhiqiang Xie, Mingyu Ding, and Ping Luo. Robotwin: Dual-arm robot benchmark with generative digital twins, 2025. URL https://arxiv.org/abs/2504.13059. [49]AgiBot-World-Contributors, Qingwen Bu, Jisong Cai, Li Chen, Xiuqi Cui, Yan Ding, Siyuan Feng, Shenyuan Gao, Xindong He, Xuan Hu, Xu Huang, Shu Jiang, Yuxin Jiang, Cheng Jing, Hongyang Li, Jialu Li, Chiming Liu, Yi Liu, Yuxiang Lu, Jianlan Luo, Ping Luo, Yao Mu, Yuehan Niu, Yixuan Pan, Jiangmiao Pang, Yu Qiao, Guanghui Ren, Cheng Ruan, Jiaqi Shan, Yongjian Shen, Chengshi Shi, Mingkang Shi, Modi Shi, Chonghao Sima, Jianheng Song, Huijie Wang, Wenhao Wang, Dafeng Wei, Chengen Xie, Guo Xu, Junchi Yan, Cunbiao Yang, Lei Yang, Shukai Yang, Maoqing Yao, Jia Zeng, Chi Zhang, Qinglin Zhang, Bin Zhao, Chengyue Zhao, Jiaqi Zhao, and Jianchao Zhu. Agibot world colosseo: A large-scale manipulation platform for scalable and intelligent embodied systems, 2025. URL https://arxiv.org/abs/2503.06669. [50]Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, Szymon Jakubczak, Tim Jones, Liyiming Ke, Sergey Levine, Adrian Li-Bell, Mohith Mothukuri, Suraj Nair, Karl Pertsch, Lucy Xiaoyang Shi, James Tanner, Quan Vuong, Anna Walling, Haohuan Wang, and Ury Zhilinsky.ํ 0 : A vision-language-action flow model for general robot control, 2026. URL https://arxiv.org/abs/2410.24164. [51]Ricardo Garcia, Shizhe Chen, and Cordelia Schmid. Towards generalizable vision-language robotic manipulation: A benchmark and llm-guided 3d policy, 2025. URLhttps://arxiv.org/abs/2410. 01345. 16 [52]Wilbert Pumacay, Ishika Singh, Jiafei Duan, Ranjay Krishna, Jesse Thomason, and Dieter Fox. The colosseum: A benchmark for evaluating generalization for robotic manipulation, 2024. URL https://arxiv.org/abs/2402.08191. [53]Aravind Rajeswaran, Vikash Kumar, Abhishek Gupta, Giulia Vezzani, John Schulman, Emanuel Todorov, and Sergey Levine. Learning complex dexterous manipulation with deep reinforcement learning and demonstrations, 2018. URL https://arxiv.org/abs/1709.10087. [54]OpenAI, Marcin Andrychowicz, Bowen Baker, Maciek Chociej, Rafal Jozefowicz, Bob McGrew, Jakub Pachocki, Arthur Petron, Matthias Plappert, Glenn Powell, Alex Ray, Jonas Schneider, Szymon Sidor, Josh Tobin, Peter Welinder, Lilian Weng, and Wojciech Zaremba. Learning dexterous in-hand manipulation, 2019. URL https://arxiv.org/abs/1808.00177. [55] Yu-Wei Chao, Wei Yang, Yu Xiang, Pavlo Molchanov, Ankur Handa, Jonathan Tremblay, Yashraj S. Narang, Karl Van Wyk, Umar Iqbal, Stan Birchfield, Jan Kautz, and Dieter Fox. Dexycb: A benchmark for capturing hand grasping of objects, 2021. URL https://arxiv.org/abs/2104.04631. [56] Jialiang Zhang, Haoran Liu, Danshi Li, Xinqiang Yu, Haoran Geng, Yufei Ding, Jiayi Chen, and He Wang. Dexgraspnet 2.0: Learning generative dexterous grasping in large-scale synthetic cluttered scenes, 2024. URL https://arxiv.org/abs/2410.23004. [57]Jianglong Ye, Keyi Wang, Chengjing Yuan, Ruihan Yang, Yiquan Li, Jiyue Zhu, Yuzhe Qin, Xueyan Zou, and Xiaolong Wang. Dex1b: Learning with 1b demonstrations for dexterous manipulation, 2025. URL https://arxiv.org/abs/2506.17198. [58]Zijie Zhu, Weiren Cai, Yizhou Wang, Zhenjie Yang, Yide Liu, Jiahao Chen, and Guanqi He. Mint: A unified model for world-space camera and hand motion estimation from scalable egocentric pipeline supervision. arXiv preprint arXiv:2609.04958, 2026. [59] Shuqi Zhao, Xinghao Zhu, Yuxin Chen, Chenran Li, Lichen Xie, Xiang Zhang, Mingyu Ding, and Masayoshi Tomizuka. Dexh2r: Task-oriented dexterous manipulation from human to robots, 2026. URL https://arxiv.org/abs/2411.04428. [60] Ryan Hoque, Peide Huang, David J. Yoon, Mouli Sivapurapu, and Jian Zhang. Egodex: Learning dexterous manipulation from large-scale egocentric video, 2026. URLhttps://arxiv.org/abs/ 2505.11709. [61] Zhenjie Yang, Xingyu Jiao, Guopeng Zhong, Shuzhe Yang, Shi Che, Chao Wu, Chenyu Jiang, Dongjie Zhang, Yideng Zhang, Zheng Zhang, et al. Handedit: A unified benchmark for egocentric human-to-robot dexterous hand image editing. arXiv preprint arXiv:2608.12122, 2026. [62] ALOHA 2 Team, Jorge Aldaco, Travis Armstrong, Robert Baruch, Jeff Bingham, Sanky Chan, Kenneth Draper, Debidatta Dwibedi, Chelsea Finn, Pete Florence, Spencer Goodrich, Wayne Gramlich, Torr Hage, Alexander Herzog, Jonathan Hoech, Thinh Nguyen, Ian Storz, Baruch Tabanpour, Leila Takayama, Jonathan Tompson, Ayzaan Wahid, Ted Wahrburg, Sichun Xu, Sergey Yaroshenko, Kevin Zakka, and Tony Z. Zhao. Aloha 2: An enhanced low-cost hardware for bimanual teleoperation, 2024. URL https://arxiv.org/abs/2405.02292. [63]Cheng Chi, Zhenjia Xu, Chuer Pan, Eric Cousineau, Benjamin Burchfiel, Siyuan Feng, Russ Tedrake, and Shuran Song. Universal manipulation interface: In-the-wild robot teaching without in-the-wild robots, 2024. URL https://arxiv.org/abs/2402.10329. [64]Zhaxizhuoma, Kehui Liu, Chuyue Guan, Zhongjie Jia, Ziniu Wu, Xin Liu, Tianyu Wang, Shuai Liang, Pengan Chen, Pingrui Zhang, Haoming Song, Delin Qu, Dong Wang, Zhigang Wang, Nieqing Cao, Yan Ding, Bin Zhao, and Xuelong Li. Fastumi: A scalable and hardware-independent universal manipulation interface with dataset, 2025. URL https://arxiv.org/abs/2409.19499. 17 [65]Wenzhen Yuan, Siyuan Dong, and Edward H. Adelson. Gelsight: High-resolution robot tactile sensors for estimating geometry and force. Sensors, 17(12), 2017. ISSN 1424-8220. doi: 10.3390/s17122762. URL https://w.mdpi.com/1424-8220/17/12/2762. [66]Stephen Tian, Frederik Ebert, Dinesh Jayaraman, Mayur Mudigonda, Chelsea Finn, Roberto Calandra, and Sergey Levine. Manipulation by feel: Touch-based control with deep predictive models, 2019. URL https://arxiv.org/abs/1903.04128. [67]Siyuan Dong, Devesh K. Jha, Diego Romeres, Sangwoon Kim, Daniel Nikovski, and Alberto Rodriguez. Tactile-rl for insertion: Generalization to objects of unknown geometry, 2021. URLhttps://arxiv. org/abs/2104.01167. [68] Zhengtong Xu and Yu She. Letac-mpc: Learning model predictive control for tactile-reactive grasping, 2024. URL https://arxiv.org/abs/2403.04934. [69]Raunaq Bhirangi, Venkatesh Pattabiraman, Enes Erciyes, Yifeng Cao, Tess Hellebrekers, and Lerrel Pinto. Anyskin: Plug-and-play skin sensing for robotic touch, 2024. URLhttps://arxiv.org/abs/ 2409.08276. [70] Carolina Higuera, Akash Sharma, Chaithanya Krishna Bodduluri, Taosha Fan, Patrick Lancaster, Mrinal Kalakrishnan, Michael Kaess, Byron Boots, Mike Lambeta, Tingfan Wu, and Mustafa Mukadam. Sparsh: Self-supervised touch representations for vision-based tactile sensing, 2024. URLhttps: //arxiv.org/abs/2410.24090. [71]Carmelo Sferrazza, Younggyo Seo, Hao Liu, Youngwoon Lee, and Pieter Abbeel. The power of the senses: Generalizable manipulation from vision and touch through masked multimodal learning, 2023. URL https://arxiv.org/abs/2311.00924. [72]Abraham George, Selam Gano, Pranav Katragadda, and Amir Barati Farimani. Vital pretraining: Visuo-tactile pretraining for tactile and non-tactile manipulation policies, 2024. URLhttps://arxiv. org/abs/2403.11898. [73]Han Xue, Jieji Ren, Wendi Chen, Gu Zhang, Yuan Fang, Guoying Gu, Huazhe Xu, and Cewu Lu. Reactive diffusion policy: Slow-fast visual-tactile policy learning for contact-rich manipulation, 2025. URL https://arxiv.org/abs/2503.02881. [74]Yansong Wu, Zongxie Chen, Fan Wu, Lingyun Chen, Liding Zhang, Zhenshan Bing, Abdalla Swikir, Sami Haddadin, and Alois Knoll. Tacdiffusion: Force-domain diffusion policy for precise tactile manipulation, 2025. URL https://arxiv.org/abs/2409.11047. [75] Peng Hao, Chaofan Zhang, Dingzhe Li, Xiaoge Cao, Xiaoshuai Hao, Shaowei Cui, and Shuo Wang. Tla: Tactile-language-action model for contact-rich manipulation, 2025. URL https://arxiv.org/ abs/2503.08548. [76]Jialei Huang, Shuo Wang, Fanqi Lin, Yihang Hu, Chuan Wen, and Yang Gao. Tactile-vla: Unlocking vision-language-action modelโs physical knowledge for tactile generalization, 2025. URLhttps: //arxiv.org/abs/2507.09160. [77]Kaidi Zhang, Heng Zhang, Zhengtong Xu, Zhiyuan Zhang, Md Rakibul Islam Prince, Xiang Li, Xiaojing Han, Yuhao Zhou, Arash Ajoudani, and Yu She. Tacvla: Contact-aware tactile fusion for robust vision-language-action manipulation, 2026. URL https://arxiv.org/abs/2603.12665. [78]Yuzhe Huang, Pei Lin, Wanlin Li, Daohan Li, Jiajun Li, Jiaming Jiang, Chenxi Xiao, and Ziyuan Jiao. Taf-vla: Tactile-force alignment in vision-language-action models for force-aware manipulation, 2026. URL https://arxiv.org/abs/2601.20321. 18 [79]Yuan Zeng, Yujia Shi, Tiao Tan, Xingting Li, Yaqi Qin, Zongqing Lu, Wenming Yang, Jing-Hao Xue, and Qingmin Liao. EgoTactile: Learning grasp pressure for everyday objects from egocentric video, 2026. URL https://arxiv.org/abs/2606.09243. [80]Jingbo He, Michael F ฬ arber, and Roberto Calandra. RCT: A robot-collected touch-vision-language dataset for tactile generalization, 2026. URL https://arxiv.org/abs/2606.31694. [81] Yuzhe Huang, Jiaping Wu, Jiaming Jiang, Hezhe Lin, Aikebaier Aierken, Yunlong Wang, Kun Cheng, Wanlin Li, Chenxi Xiao, Ziyuan Jiao, and Yuanxin Zhong. Ht-bench: Benchmarking and learning dexterous full-hand tactile representations with egocentric vision, 2026. URLhttps: //arxiv.org/abs/2606.19161. [82] Jongbin Lim, Taeyun Ha, Mingi Choi, Jisoo Kim, Byungjun Kim, Subin Jeon, and Hanbyul Joo. HRDexDB: A paired human-robot dataset for cross-embodiment dexterous grasping, 2026. URL https://arxiv.org/abs/2604.14944. [83]Elle Miller, Jayaram Reddy, Ayush Deshmukh, Trevor McInroe, David Abel, Oisin Mac Aodha, and Sethu Vijayakumar. roto 2.0: The robot tactile olympiad, 2026. URLhttps://arxiv.org/abs/ 2605.21429. [84] Bowen Jing, Mingxin Wang, Ruiyang Hao, Chenchen Ge, Hanwen Shen, Junjie He, Yang Cui, Yiming Hou, Weitao Zhou, Jiawei Wang, Minglei Li, Dandan Zhang, Ding Zhao, Houde Liu, Xiaofan Li, Si Liu, Ping Luo, and Haibao Yu. Softvtbench: A safety-aware visuo-tactile benchmark for physically constrained robotic manipulation of deformable objects (early version), 2026. URL https://arxiv.org/abs/2607.04234. [85]Lei Su, Zhijie Peng, Renyuan Ren, Shengping Mao, Juan Du, Kaifeng Zhang, and Xuezhou Zhu. Tacmap: Bridging the tactile sim-to-real gap via geometry-consistent penetration depth map, 2026. URL https://arxiv.org/abs/2602.21625. [86]Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion, 2024. URL https://arxiv.org/abs/2303.04137. [87]Physical Intelligence, Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Manuel Y. Galliker, Dibya Ghosh, Lachy Groom, Karol Hausman, Brian Ichter, Szymon Jakubczak, Tim Jones, Liyiming Ke, Devin LeBlanc, Sergey Levine, Adrian Li-Bell, Mohith Mothukuri, Suraj Nair, Karl Pertsch, Allen Z. Ren, Lucy Xiaoyang Shi, Laura Smith, Jost Tobias Springenberg, Kyle Stachowicz, James Tanner, Quan Vuong, Homer Walke, Anna Walling, Haohuan Wang, Lili Yu, and Ury Zhilinsky.ํ 0.5 : a vision-language-action model with open-world generalization, 2025. URL https://arxiv.org/abs/2504.16054. [88]NVIDIA, :, Johan Bjorck, Fernando Casta ฬ neda, Nikita Cherniadev, Xingye Da, Runyu Ding, Linxi โJimโ Fan, Yu Fang, Dieter Fox, Fengyuan Hu, Spencer Huang, Joel Jang, Zhenyu Jiang, Jan Kautz, Kaushil Kundalia, Lawrence Lao, Zhiqi Li, Zongyu Lin, Kevin Lin, Guilin Liu, Edith Llontop, Loic Magne, Ajay Mandlekar, Avnish Narayan, Soroush Nasiriany, Scott Reed, You Liang Tan, Guanzhi Wang, Zu Wang, Jing Wang, Qi Wang, Jiannan Xiang, Yuqi Xie, Yinzhen Xu, Zhenjia Xu, Seonghyeon Ye, Zhiding Yu, Ao Zhang, Hao Zhang, Yizhou Zhao, Ruijie Zheng, and Yuke Zhu. Gr00t n1: An open foundation model for generalist humanoid robots, 2025. URL https://arxiv.org/abs/2503.14734. 19 Supplementary Materials A Full Task Catalog Table S1 lists the tasks. For each task we report the task identifier and name, a short English description, the success condition used by the evaluator, the robot embodiment, and a scene snapshot. Table S1. Full Bench2Dex task catalog: identifier, description, success condition, robot embodiment, and scene snapshot. TaskDescriptionSuccess ConditionEmbodimentScene Wine Glass Plate Balance Carry three filled wine glasses to the plates of three diners without spilling or knocking over any objects. All three wine glasses are kept upright (within 30 โฆ of vertical), each placed within 0.10 m of one of the target positions, and fully at rest. xArm7+ Ability Fruit Bowl Loading Move the bowl to the center of the table and place two apples and a banana into it. The bowl is moved into the central zone of the table and kept upright, and two apples and one banana are placed inside it, all at rest. UR5+ RH56DFX Citrus Plate Loading Place two lemons and two oranges onto the plate. Two lemons and two oranges are placed on the plate, with each objectโs center within0.15m of the plate interior. UR5+ RH5DG2 Frypan Stand & Pour Place the frypan onto the display stand, then pour the soy sauce and olive oil into the frypan one by one. The frypan is placed on the stand; the soy sauce and olive oil are each tilted at least 50 โฆ with their mouth over the frypan to pour, then both bottles are returned upright and at rest on the table while the bread remains in the pan. UR5+Schunk Cleaner Box Loading Use both hands to lift the cleaner upright into the wooden box, then place the soap into the box. The wooden box is upright, the cleaner is inside the box and upright (not tilted), and the soap is inside the box. xArm7+ LEAP Screwdriver Box & Hammer Place the right screwdriver into the box and move the box left, place the left screwdriver into the box, then strike the wooden block once with the hammer and place the hammer into the box. Both screwdrivers are inside the upright box, the hammer strikes the wooden block once (swing near the block with a rigid-body response), and the hammer is then placed inside the box. UR5+ RH56DFX Condiment Box Loading Place the monosodium glutamate, soy sauce, and vinegar into the wooden box. The box is upright; the MSG, soy sauce, and vinegar are all inside the box, with the soy sauce and vinegar kept upright, and all objects at rest. UR5+Wuji Tool Box Loading Use the right hand to place the drill and flat screwdriver and the left hand to place the wrench and Phillips screwdriver into the wooden box. All four tools (drill, flat screwdriver, adjustable wrench, Phillips screwdriver) are placed inside the upright wooden box. Panda+Orca Stationery Category Sorting Use the right hand to sort the marker and battery into the pen cup and plastic box, and the left hand to sort the gray pen and glue into the respective containers. The pen and marker are placed upright in the pen cup and the glue and battery are placed in the plastic box, with both containers upright and all objects at rest. RM65+Revo2 Continued on next page 20 Table S1. Full Bench2Dex task catalog (continued). TaskDescriptionSuccess ConditionEmbodimentScene Canned Food Tray Arrange- ment Use the left hand to place the master chef can and MSG and the right hand to place the milk box and potted meat can onto the tray. The master chef can, MSG, milk box, and potted meat can are all placed on the tray and kept upright, with the tray upright and all objects at rest. IIWA7+ Sharpa Ball Box Loading Place the mini soccer ball, tennis ball, golf ball, and ping-pong ball into the box. The box is upright and all four balls are placed inside it, each within 0.22 m of the box interior. UR5+Wuji Baking Tray Prep Use the left hand to place the brush and spatula and the right hand to place the small pudding box and large gelatin box onto the tray. The brush, spatula, pudding box, and gelatin box are all placed on the tray, with the tray kept upright and all objects at rest. IIWA7+ Sharpa Fridge Wine Interhand Pour Open the refrigerator, take out the wine bottle, hand it to the right hand in the air to pour a glass of wine, return it to the left hand to put back into the fridge, and close the door. The fridge door is opened, the wine bottle is lifted out, tilted at least 50 โฆ toward the glass to pour, then returned upright inside the fridge, and the door is closed. UR5+ RH5DG2 Trash Disposal Open the trash can lid, throw the crumpled paper into the trash can, use the left hand to throw the bottle and banana peel into the trash can, then close the lid. The trash can is upright and the crumpled paper, bottle, and banana peel are all placed inside it, at rest. UR5+ RH56DFX Fridge Fruit Shelf Sorting Move the banana to the left side of the desk, open both fridge doors, place the lemon on the upper shelf and the banana on the lower shelf, then close the door. The banana and lemon are placed inside the fridge, both fridge doors are closed, and both objects are at rest. UR5+Shadow Microwave Bowl Loading Open the microwave door, bring the bowl to the microwave, place the baguette into the bowl, place the bowl into the microwave, and close the door. The bowl is inside the microwave cavity and upright, the baguette is placed in the bowl, and the microwave door is closed. UR5+Schunk Toilet Lid Cleaner Pour Open the toilet lid, pour the cleaner into the toilet, and close the toilet lid. The toilet lid is opened, the cleaner is lifted and tilted at least 50 โฆ toward the toilet bowl to pour while the lid stays open, and the lid is then closed. xArm7+ LEAP Breadbasket Fast-Food Loading Place the baguette and bread into the bread basket, move the basket to the center of the table, then place the hamburger and french fries into the basket. The bread basket is upright and the french fries, hamburger, bread, and baguette are all placed inside it. UR5+ RH5DG2 Medicine Shoebox Pack Move the shoe box to the center of the table, then place the pill bottle, toothpaste, and hydrating oil into the box. The shoe box is upright and the pill bottle, toothpaste, and hydrating oil are all placed inside it. Panda+Orca Shoebox Accessory Pack Place the seal into the shoe box and move the box to the center of the table, then place the shoe and pet collar into the box. The shoe box is upright and the seal, shoe, and pet collar are all placed inside it, at rest. xArm7+ LEAP Continued on next page 21 Table S1. Full Bench2Dex task catalog (continued). TaskDescriptionSuccess ConditionEmbodimentScene Sports Ball Cup Sort Place the tennis ball and baseball into the large cup and the racquetball and golf ball into the small cup. Both cups are upright; the tennis ball and baseball are inside the large cup and the racquetball and golf ball are inside the small cup, all at rest. Panda+ Allegro Faucet Cup Water Fill Place the spoon into the mug, place the mug under the faucet, open the faucet to fill the mug and then close it, and place the mug onto the tray. The spoon is placed in the upright mug; while the mug remains under the faucet opening, the faucet is opened and then closed again; and the mug is placed at rest on the tray. RM65+Revo2 Jigsaw Puzzle Assembly Use the right and left hands in turn to assemble the green, red, blue, and yellow puzzle pieces onto the fixed white puzzle piece in the center of the table, forming a rectangle. All four colored puzzle pieces are placed at their target positions around the fixed white piece (within0.02m inํฅํฆand0.01m in height of the reference), and all pieces are at rest. IIWA7+ Sharpa Soup Serving Hold the bowl beside the pot with the left hand, ladle two scoops of soup from the pot into the bowl with the right hand, carry the bowl to the front-right area, and return the ladle. The pot is on the stove, the bowl is held near the pot, the ladle dips into the pot and is brought over the bowl twice, the bowl is delivered to the front-right serving zone, and the ladle is returned. UR5+Shadow Bimanual Piano Melody Both hands play the piano key sequence C-C-G-G-A-A-G together, the left hand playing the bass notes and the right hand the treble notes. The left hand plays the bass melody and the right hand plays the treble melody C-C-G-G-A-A-G, with each note key pressed in sequence. xArm7+ Ability Gaming Desk Setup Straighten the monitor screen forward with both hands, press the ESC key, move the mouse back to the left side of the keyboard, and click the left mouse button. The monitor is straightened forward, the ESC key is pressed once, the mouse is returned to the left of the keyboard, and the left mouse button is clicked. JAKA ZU7+ DexHand021 B Supported Bimanual Robot Embodiments This section summarizes the 12 bimanual robot embodiments supported by Bench2Dex. Each embodiment combines a robot arm and a dexterous hand, enabling evaluation across different arm kinematics, hand morphologies, palm geometries, and finger configurations. Table S2. Supported bimanual robot embodiments in Bench2Dex. EmbodimentRenderDescription JAKA ZU7+DexHand021 A metallic-gray collaborative arm with blue circular joint caps, paired with a silver dexterous hand with dense exposed mechanisms, a ribbed palm surface, and slender articulated fingers. The transparent wrist housing and mechanical finger structure make it distinctive. IIWA7+Sharpa A white KUKA-style arm with bright orange accent panels, paired with a smooth anthropomorphic hand. The clean enclosed hand silhouette contrasts with the strong white-orange arm styling. Continued on next page 22 Table S2. Supported bimanual robot embodiments in Bench2Dex (continued). EmbodimentRenderDescription Panda+Allegro A light Panda-style arm paired with a compact dark modular hand. The boxy palm, rectangular segmented fingers, and bright fingertip caps create a sharp contrast with the soft industrial arm. Panda+Orca A light Panda-style arm attached to a dark hand with a narrow palm and strongly spread fingers. Long separated digits with bright caps create an open fan-like silhouette. RM65+Revo2 A smooth white arm paired with a compact anthropomorphic hand with a rounded palm shell, slim fingers, and a clean enclosed structure. The embodiment appears polished and tightly integrated. UR5+RH56DFX A metallic-gray industrial arm with rounded cylindrical links and blue joint caps, ending in a white-and-silver dexterous hand with a sculpted palm shell, dark finger pads, and a large side thumb. UR5+RH5DG2 A metallic-gray UR5-style arm with blue circular joint caps, attached to a white-and-gray hand with a rounded palm shell, dark cylindrical finger coverings, and a thick side thumb close to the palm plane. UR5+Schunk A metallic-gray arm with blue round joint covers, paired with a robust white industrial hand. A broad palm, thick segmented fingers, gray fingertip pads, and heavy thumb create a sturdy precision-oriented look. UR5+Shadow A metallic-gray arm with blue circular joints, attached to a dark anthropomorphic hand with slim multi-joint fingers and bright fingertip caps. The hand is lightweight and human-like relative to the heavier arm. UR5+Wuji A metallic-gray arm with blue circular joint caps, ending in a slim dexterous hand with long narrow fingers and a thin side thumb. The hand appears lighter and more elongated than other UR5-based embodiments. xArm7+Ability A white arm with smooth enclosed links, paired with a compact white hand with a rounded palm shell, dark finger coverings, and a thick side thumb. The embodiment is soft-contoured and highly enclosed. xArm7+LEAP A white arm with smooth rounded links, attached to a compact low-profile robotic hand with a blocky palm and modular fingers. Stacked box-like segments and bright caps give it a clean geometric appearance. C Unified Tactile Surface Reconstruction Bench2Dex reconstructs tactile contact surfaces for the twelve bimanual hand embodiments included in the benchmark. The representation builds on image-based tactile simulation [27,28] and the geometry-consistent penetration-depth formulation of TacMap [85], while extending the surface-map abstraction to heterogeneous hand morphologies. For each hand, we identify the physical regions that should generate tactile evidenceโ elastomer pads, rubber pads, fingertip shells, force-sensor covers, or dedicated touch linksโand rebuild or clean them in Blender. The cleaned meshes are then converted into surface point-and-normal assets so 23 that the tactile signal is defined on the intended contact surface rather than on arbitrary full-link geometry. The reconstruction follows a common convention across embodiments. Each surface mesh is expressed in the tactile sensor attach-link frame, or converted from the original URDF visual or collision geometry using the corresponding geometry origin, rotation, and mesh scale. The mesh is rasterized into a regular surface-aligned tactile grid, producing point and inward-facing normal maps at the native resolution used by the tactile registry, so that runtime rays are cast from the nominal contact surface toward the finger interior. The generated points are stored in millimeters and converted back to meters during sensing, keeping offline asset generation and runtime ray-casting consistent. Different hands expose different mesh sharing patterns: some use a shared non-thumb surface for all non-thumb fingers, while others require independent assets per finger because the pad geometry varies. Table S3 summarizes the resulting grouping policy. The current 12-hand setup uses a four-finger-shared grouping for nine hands and a per-finger independent grouping for three hands. Each hand places one tactile site on every finger, so the ten five-finger embodiments carry ten sites each, while the four-finger Allegro and LEAP embodiments carry eight; palm sites are excluded. In the released data, each site is sampled at the native surface-map resolution of240ร 240(subsampling step 1) and captured once per simulation step during replay. Table S3. Tactile surface-map grouping policy for the twelve Bench2Dex hands. โ4Fโ denotes a shared non-thumb surface group. Grouping policyHandsSurface-map asset groups Legacy side-shared thumb/4F SharpaTwo assets are shared across both hands: TH for thumbs and 4F for all non-thumb fingers. This preserves the original Sharpa TacMap asset convention. Side-specific thumb/4F shared RH5DG2, Shadow, Schunk, Wuji, Allegro, Orca, LEAP, DexHand021 Four assets are used per embodiment: RTH, R4F, LTH, and L4F. The non-thumb group uses one representative surface per side. Allegro and LEAP are four-finger hands, so 4F covers index, middle, and ring. Per-finger independentRH56DFX, Ability, Revo2Each finger has its own surface-map group on each side, such as RTH, RIDX, RMID, RRING, RLIT and the corresponding left-hand groups. This is used when non-thumb pad meshes are not safely interchangeable. Table S5. Tactile contact-surface mesh checklist used for Blender reconstruction. The mesh column lists only source filenames rather than full dataset paths. HandGroupingSource links for reconstructionMesh source used in BlenderNotes DexHand021Side-specific thumb/4F Thumb: rflink14 and lflink14. 4F representative: rflink24 and lflink24. rflink14.STL; lflink14.STL; rflink24.STL; lflink24.STL. The source is the fourth link of each finger. The current 4F group uses finger 2 as the representative non-thumb surface. SharpaLegacy side-shared thumb/4F Thumb representative: rightthumbelastomer. 4F representative: rightindexelastomer. thumbelastomersurface.STL; elastomersurface.STL. Preserves the original TH/4F naming. The legacy assets store outward-facing normals, which are flipped at runtime so that all hands share the inward-facing normal convention. Continued on next page 24 Table S5. Tactile contact-surface mesh checklist used for Blender reconstruction (continued). HandGroupingSource links for reconstructionMesh source used in BlenderNotes AllegroSide-specific thumb/4F Thumb: multilink15.0tip and link15.0tip. 4F representative: multi link3.0tip and link 3.0tip. linktip.obj.All fingertips reuse the same tip mesh. The Blender export keeps the fingertip offset convention, with side-specific sensor normals. OrcaSide-specific thumb/4F Thumb: multirightthumbdp and leftthumbdp. 4F representative: multirightindexip and left indexip. rightcollisionthumbdpskinmesh.stl; left collisionthumbdpskinmesh.stl; rightcollisionindexipskinmesh.stl; left collisionindexipskinmesh.stl. Uses collision skin geometry rather than visual body geometry. Left and right origins/rotations differ, so side-specific meshes are required. Revo2Per-finger independent Right and left thumb touchlink, indextouchlink, middle touchlink, ringtouchlink, and pinkytouchlink. rightthumbtouchlink.STL; right indextouchlink.STL; rightmiddletouchlink.STL; right ringtouchlink.STL; rightpinkytouchlink.STL; left-hand equivalents. Each touch link is a tactile surface. The non-thumb meshes differ enough that a shared 4F map would lose edge fidelity. RH56DFXPer-finger independent Right and left thumbrubber3, indexrubber2, middle rubber2, ringrubber2, and littlerubber2. rightthumbrubber3.STL; rightindexrubber2.STL; right middlerubber2.STL; rightringrubber2.STL; rightlittlerubber2.STL; left-hand equivalents. Rubber pad dimensions differ across fingers, especially the little finger, so each finger is reconstructed separately. RH5DG2Side-specific thumb/4F Thumb: right thumbforcesensor and leftthumbforcesensor. 4F representative: right indexforcesensor and leftindexforcesensor. rightthumbforcesensor.STL; leftthumbforcesensor.STL; rightindexforcesensor.STL; leftindexforcesensor.STL. The force-sensor meshes are used as the contact pads. The non-thumb force-sensor geometry is treated as shareable within each side. SchunkSide-specific thumb/4F Thumb: right handc and lefthandc. 4F representative: righthandt and lefthandt. d13.obj; d13 left.obj; fingertip.obj.Site-body overrides move the tactile sites from distal bodies to fingertip bodies. Non-thumb fingers reuse the same fingertip mesh. ShadowSide-specific thumb/4F Thumb: thdistal and lthdistal. 4F representative: ffdistal and l ffdistal. th distalpst.obj; fdistalpst.obj.The Shadow URDF applies a 0.001 mesh scale. The non-thumb distal mesh is shared across the four non-thumb fingers. WujiSide-specific thumb/4F Thumb: right finger1tiplink and leftfinger1tiplink. 4F representative: right finger2tiplink and leftfinger2tiplink. rightfinger1tiplink.STL; leftfinger1tiplink.STL; right finger2tiplink.STL; leftfinger2tiplink.STL. Tactile sites attach to the true fingertip links. Finger 2 represents the non-thumb group in the 4F-shared version. AbilityPer-finger independent Right and left thumbL2, indexL2, middleL2, ringL2, and pinkyL2. thumb F2right.STL (right thumb); thumbF2left.STL (left thumb); idx F2Lg.STL (index, middle, ring, and left pinky); idxF2.STL (right pinky). The right-hand pinky source mesh is smaller than the shared non-thumb mesh, so Ability uses per-finger surface-map groups. LEAPSide-specific thumb/4F Thumb: thumbfingertip. 4F representative: fingertip. thumbfingertip.obj; fingertip.obj.LEAP is a four-finger hand. The non-thumb fingertip mesh is shared, but its URDF visual origin must be preserved unless the Blender export is already in the attach-link frame. Table S5 lists the source meshes used for Blender reconstruction, and Table S6 visualizes the resulting unified tactile surfaces for all twelve hands. 25 Table S6. Reconstructed tactile contact surfaces for the twelve Bench2Dex hands. HandReconstructed tactile surface DexHand021 Sharpa Allegro Orca Revo2 Continued on next page 26 Table S6. Reconstructed tactile contact surfaces for the twelve Bench2Dex hands (continued). HandReconstructed tactile surface RH56DFX RH5DG2 Schunk Shadow Wuji Continued on next page 27 Table S6. Reconstructed tactile contact surfaces for the twelve Bench2Dex hands (continued). HandReconstructed tactile surface Ability LEAP C.1 Signal Semantics and Depth Quantization The raw value before quantization is the ray-cast distance along the inward-facing local surface normal: each ray is launched from a nominal surface point toward the finger interior, so the first-hit distance measures the depth to which an object surface has penetrated past the nominal contact surface, rather than the gap to an approaching but non-contacting object. Letํ ํ ,ํก (ํข,ํฃ)denote this metric distance, in meters, at tactile pixel (ํข,ํฃ)for siteํ and frameํก. The contact-consistency check is a two-pass test: after the first hit, the ray origin is advanced to just short of the hit point and a reverse ray is cast; the sample is kept only when the reverse ray also intersects the target mesh, which holds once the object surface has crossed the rayโs nominal surface origin. Rays are cast against the meshes of the dynamic task objects and, for articulated task objects, against their mesh-bearing rigid links, within a maximum cast range of15m; static scene geometry such as the tabletop is not among the raycast targets and produces no tactile response. Missed rays (no target hit within this range) and samples that fail the check are assigned zero distance and are additionally recorded by a binary contact-validity mask. For valid samples, the distance is first converted to millimeters, ํง ํ ,ํก (ํข,ํฃ) = 10 3 ํ ํ ,ํก (ํข,ํฃ),(4) and then mapped to an 8-bit integer through the piecewise quantizer used by the released TacMap implemen- tation [85], ํ(ํง) = ( ํง/0.005,ํง < 0.5, (ํงโ 0.5)/0.03+ 100, ํง โฅ 0.5, (5) followed by clipping and casting, e T ํ ,ํก (ํข,ํฃ) = uint8 clip ํ(ํง ํ ,ํก (ํข,ํฃ)), 0, 255 .(6) The quantized map is then spatially smoothed with a Gaussian kernel evaluated in floating point. The filtered response is rounded and cast back to an 8-bit integer before being stored in the tactile stream, T ํ ,ํก = uint8 round G ํ โ e T ํ ,ํก ,(7) 28 Table S7. Eight data modalities recorded by Bench2Dex. ModalityDescription RGB imagesMulti-camera visual observations for policy learning and video analysis. Depth maps Geometric observations for 3D perception and reconstruction. Robot statesJoint positions, velocities, efforts, and proprioceptive state information. Object statesObject poses, velocities, and task-relevant physical states. Tactile signalsUnified surface-aligned tactile images defined on robot hand contact surfaces. 2D boxesImage-space object annotations for visual detection supervision. 3D boxes Object-level 3D bounding boxes for geometric supervision. OccupancyTSDF or mesh-based occupancy labels for 3D scene understanding. whereG ํ denotes Gaussian filtering and the current implementation usesํ = 1.5; the kernel size is max(1,โ9/stepโ)rounded up to an odd integer, which gives9ร 9at the released subsampling step of 1. The first branch allocates finer precision to small contact depths, using0.005m per integer level forํง < 0.5 m, while the second branch uses a coarser0.03m per level for larger geometric penetration-depth values. Values of5.15m and above saturate at 255. Thus, each tactile-map pixel preserves high sensitivity near initial contact while remaining compact as a single 8-bit value in the HDF5 tactile stream. D HDF5 Schema Details Table S8. Full per-episode HDF5 organization used by Bench2Dex when all eight modalities are collected.ํdenotes the number of recorded frames,ํฝthe robot joint dimension,ํปรํthe camera resolution, andํป ํก รํ ํก the surface-aligned tactile resolution (240ร 240 in the released data). HDF5 nodeStored content episodeid.hdf5One synchronized trajectory episode. The parent directory also stores an episode manifest.json index. meta/Task, scene, robot key, modality list, camera definitions, collection config, instruction, success flag, schema version, and dataset metadata. time/ frameindex, timestampns, and simstep for temporal alignment across all modalities. episode/Episode flags: isfirst, islast, done, and success. action/Commanded robot action, action validity, action source, joint/action names, control mode, and action type. framevalid, frameerrorsPer-frame validity flags and error messages for filtering corrupted or partially captured frames. cameras/cameraid/Per-camera visual observations and calibration metadata. rgbRGB images, shape[ํ, ํป,ํ,3]when stored as dense images. Modality: RGB images. depthm, depthsemanticsMetric depth maps and depth convention, shape [ํ, ํป,ํ]. Modality: depth maps. intrinsic, extrinsic worldfromcamCamera intrinsics and per-frame world-from-camera extrinsics. cameramodel, fisheye calibrationCamera model, clipping range, and optional fisheye matrix, distortion coefficients, and valid mask. robot/Robot proprioception and tactile observations. joint names, qpos, qvel, qeffortJoint names and per-frame joint position, velocity, and effort, shape [ํ, ํฝ]. Modality: robot states. tactile/metaSurface-aligned tactile metadata, including the tactile site list and sensor settings. tactile/tacmap/sitenameQuantized surface-aligned tactile images, shape [ํ, ํป ํก ,ํ ํก ], uint8. Modality: tactile signals. tactile/distance alongnormalmGroup of per-site raw metric ray-depth datasets (float32, meters), one dataset per tactile site, each of shape [ํ, ํป ํก ,ํ ํก ]. tactile/contactmaskGroup of per-site binary contact-validity datasets (bool), one dataset per tactile site, each of shape [ํ, ํป ํก ,ํ ํก ]; true where the sample passed the contact-consistency check. objects/objid/Per-object physical state traces. pose worldObject pose in world frame, shape [ํ, 7] for position and quaternion. Modality: object states. linvelworld, angvelworldObject linear and angular velocity traces, shape [ํ, 3]. joint names, qpos, qvel, qeffortOptional articulated-object joint state fields. labels/box2d/cameraid/objid/Image-space boxes xyxy and visibility flags, shapes [ํ, 4] and [ํ]. Modality: 2D boxes. labels/box3d/obj id/Oriented 3D boxes: centerworld, sizelwh, and quatworld. Modality: 3D boxes. labels/occupancy tsdf/Occupancy grid state, grid shape, bounds, voxel size, method, semantics, and frame-valid flags. Modality: occupancy. metrics/episodeEpisode-level task progress, success, efficiency, and safety metrics when available. metrics/timeseriesPer-frame or per-step metric traces aligned with the episode timeline. E Evaluation Metrics Bench2Dex uses a unified metric suite that separates core benchmark scores from auxiliary diagnostic signals. Each evaluated task defines executable terminal conditions, stage-level subgoals, safety rules, and, when applicable, grasp or tool-use diagnostics. Letํ,ํ, andํindex the task, policy, and evaluation channel, respectively, and letํdenote the number of recorded episodes in a(ํ,ํ, ํ)cell unless stated otherwise. 29 Episode-level quantities are first summarized within each cell, and benchmark-level episode metrics are equal-weight means across tasks. For metrics defined as means over allํepisodes, this task-macro mean equals the corresponding pooled episode mean because every reported cell contains 50 episodes. Unless noted otherwise, the benchmark uses the reach-and-stop protocol; fixed-horizon evaluation is treated as a distinct protocol and identified explicitly in the evaluation summary. E.1 Primary Metrics Metric set. The core reported metric set is M core =SR, LSCR,ํก succ , SafeSR, HVR, DropR, HSR, VSR step , R.(8) SR is the primary completion outcome; LSCR and SafeSR provide complementary views of progress and safety, while the remaining quantities characterize success-conditional efficiency, violations, and robustness. SR can be accompanied by a central 95% confidence interval computed from 10,000 percentile-bootstrap resamples of the corresponding episode group with a fixed seed of 0. Reach-and-stop success. In the reported evaluations, the evaluator is updated once per executed physics step. Letํถ ํ,ํ โ 0, 1be the terminal predicate at evaluator updateํ,ํฟํก ํ,ํ the elapsed simulation time since the preceding update,ํ ํ the number of executed updates, andฮ ํ the configured dwell time. The accumulated dwell time is โ ํ,0 = 0,โ ํ,ํ = ํถ ํ,ํ โ ํ,ํโ1 + ํฟํก ํ,ํ .(9) Defineํ stable ํ = minํ โค ํ ํ : โ ํ,ํ โฅ ฮ ํ when this set is nonempty,ํ ํ = 1[ํ stable ํ exists], andํก stable ํ = ร ํ stable ํ ํ=1 ํฟํก ํ,ํ . The default dwell time is0.5s. Under reach-and-stop, an episode terminates at stable success; otherwise it ends at the evaluation horizon or when evaluation cannot proceed. A recorded episode that terminates because of a policy or runtime evaluation error remains in the denominator and is counted as unsuccessful. The stable success rate is SR = 1 ํ ํ โ๏ธ ํ=1 ํ ํ .(10) This protocol avoids counting transient contacts or unstable placements as successful episodes. Latched stage progress.Binary success is insufficient for long-horizon manipulation because a policy may complete early stages but fail later. Each task specifies a set of stagesG ํ =ํ ํ,ํ ํพ ํ ํ=1 with optional dependency edges. Letํฟ ํ,ํ be one if the predicate of stageํis true at some updateํ โค ํ ํ and all dependencies were latched earlier or become valid in the same-step fixed-point closure over the taskโs acyclic dependency graph; otherwise let it be zero. Once set, ํฟ ํ,ํ remains one. The primary progress metric is LSCR ํ = 1 ํพ ํ ํพ ํ โ๏ธ ํ=1 ํฟ ํ,ํ .(11) The benchmark defines aggregate LSCR as the mean ofLSCR ํ over all evaluated episodes. It measures ever-reached, dependency-valid milestones rather than final-state satisfaction; terminal-state stage status is retained as a diagnostic below. 30 Completion time. Completion time is reported as the success-conditional mean ํก succ = ร ํ ํ=1 ํ ํ ํก stable ํ ร ํ ํ=1 ํ ํ ,(12) when at least one episode succeeds; otherwise it is undefined. Because this quantity conditions on success, it should be interpreted together with SR and the number of successful episodes. Expert-normalized speed and step-based execution counts are auxiliary diagnostics rather than primary ranking fields. Safety.Safety is evaluated over executed physics steps. Letํท ํ andํป ํ denote whether episodeํcontains a task-defined drop or tracked-object high-speed violation, respectively. The hard-violation indicator is ํต ํ = ํท ํ โจ ํป ํ by default; a joint-limit rule is included only when the task explicitly promotes it to the core safety criterion. The safe-success indicator is ํ ํ (1โ ํต ํ ). Accordingly, SafeSR = 1 ํ ํ โ๏ธ ํ=1 ํ ํ (1โ ํต ํ ),(13) HVR = 1 ํ ํ โ๏ธ ํ=1 ํต ํ .(14) The associated rates areDropR = ํ โ1 ร ํ ํท ํ andHSR = ํ โ1 ร ํ ํป ํ . These events are not mutually exclusive, so their rates need not sum to HVR. Ifํ ํ,ํ indicates any task-level violation at executed stepํ, the pooled violation-step rate is VSR step = ร ํ ร ํ ํ ํ=1 ํ ํ,ํ ร ํ ํ ํ .(15) This is an exposure-normalized rate over executed steps and is interpreted together with episode-level HVR. High-speed events are severe-motion proxies, not contact-force or collision measurements. Joint-limit observations are treated as auxiliary diagnostics unless a task explicitly defines them as safety violations. Robustness.For generalized evaluation, letPcomprise the None, Equi., Inv., and Full evaluation channels. The benchmark provides each channelโs SR, confidence interval, and sample count. For a shifted channel ํ โ P\None, the relative ratio is R ํ = SR ํ SR None ,(16) whenSR None > 0. LetP + = P\Noneandํ ํ be the number of episodes in channelํ. The aggregate used by the benchmark is SR shift = ร ํโP + ํ ํ SR ํ ร ํโP + ํ ํ ,R shift = SR shift SR None .(17) R is reported together with the absolute channel-wise SRs as a supporting robustness descriptor. These are grouped, unpaired comparisons rather than paired causal estimates. E.2 Diagnostic Metrics Diagnostic metrics are not used as primary ranking fields. They expose failure modes and execution characteristics that are not uniformly applicable or comparable across all tasks and embodiments. A quantity that is undefined for a task is marked as unavailable rather than assigned a value of zero. 31 Completion diagnostics.Letํผ ํ = max ํโคํ ํ ํถ ํ,ํ indicate whether the terminal predicate is satisfied at least once, without imposing the dwell-time requirement, and letํน ํ = ํถ ํ,ํ ํ indicate whether it is satisfied at the rolloutโs actual termination observation. The corresponding ever-instantaneous and terminal-state success rates are ISR ever = 1 ํ ํ โ๏ธ ํ=1 ํผ ํ ,SR final = 1 ํ ํ โ๏ธ ํ=1 ํน ํ .(18) For stage-level diagnosis, letํด ํ,ํ (ํ ํ )indicate that the dependencies of stageํhave been satisfied by termination. The terminal-state stage completion rate is CSCR ํ = 1 ํพ ํ ํพ ํ โ๏ธ ํ=1 1 ํ ํ,ํ (ํ ํ ) = 1โง ํด ํ,ํ (ํ ํ ) = 1 .(19) Current and latched dependency-chain depths further distinguish terminal-state progress from progress reached at any earlier point. LSCR remains the primary stage-completion measure. Efficiency and execution diagnostics. For a successful episode from taskํwith a configured expert reference, task efficiency is the benchmark-specific speed ratio TE ํ = ํก expert ํ ํก stable ํ , ํก expert ํ = ํป expert ํ ํฟํก ํ ,(20) whereํป expert ํ is the task-level reference step count andํฟํก ํ is its simulation integration interval. TE is unavailable for failures or tasks without a reference. This ratio is unbounded and is comparable only under matched simulation timing, initial-state distribution, terminal condition, dwell time, and evaluation horizon. Additional execution measures include the numbers of physics steps, policy-rate control steps, and policy queries required to reach stable success; the latter two differ for chunked policies. Safety diagnostics. Safety diagnostics include violation-event density, per-type violation counts, hard joint-limit observations, finger-joint saturation, active-step rates, and maximum observed joint-limit excess. These quantities characterize kinematic constraint violations and control behavior, and they remain separate from the primary task-safety definition unless included in a taskโs safety criterion. Grasp diagnostics. For contact-rich tasks, Bench2Dex defines a kinematic grasp stability index (GSI). Since reliable handโobject contact geometry and hand forward kinematics are not uniformly available for every supported embodiment, GSI is a proxy based on object lift, stable holding, object motion, hold duration, and slip events. For a configured tracked object ํ โ O grasp ํ at step ํ , the canonical score is ํบ ํ,ํ = ํฟ ํ,ํ ํป ํ,ํ ํ ํ,ํ ํ ํ,ํ ํด ํ,ํ ,(21) where all components lie in[0, 1]. Specifically,ํฟindicates lift relative to the initial object height;ํปtakes values 1,0.35,0.15, or0for held, lifted-and-stable, lifted-only, or other states;ํ = exp[โ 1 2 (โฅํฃโฅ/ํ ํฃ +โฅํโฅ/ํ ํ )]; ํ = 0.75+ 0.25 min(1, ํ/ํ min )when held and1otherwise; andํด = 1โ min(1, ํ /ํ max ). The thresholds and scales follow the task configuration or benchmark defaults. The episode-level grasp diagnostics are GSI ํ = max ํโO grasp ํ max ํโคํ ํ ํบ ํ,ํ ,(22) GSI ํ = mean ํ,ํ:ํบ ํ,ํ >0 ํบ ํ,ํ .(23) The former records peak grasp quality, whereas the latter summarizes positive-score objectโstep observations. 32 Tool-use and motion diagnostics. Tool-use diagnostics are enabled only when stages specify tool IDs or tool equivalence classes. For episodeํ, tool selection accuracy measures whether the first held object selected for an eligible stage belongs to its allowed tool set: TSA ํ = # correct stage tool selections # tool annotated stages .(24) An annotated stage with no selected tool is counted as incorrect. Tool switch success rate measures whether an attempted switch releases the previous tool stably and grasps the next tool stably before timeout: TSSR ํ = # successful tool switches # attempted tool switches .(25) TSSR is unavailable for episodes without a switch attempt. Aggregate TSA and TSSR are means over episodes for which the corresponding quantity is defined, and TSSR is accompanied by total attempted and successful switch counts. Robot-motion diagnostics summarize joint velocity, acceleration, jerk, and effort by first computing the root mean square over available steps and joints within each episode and then averaging episode values. They are interpreted only under a shared embodiment and timing configuration. Execution summaries additionally provide the mean rollout length and the number of episodes that terminate because evaluation cannot proceed. E.3 Reproducibility Metadata Each evaluation summary is accompanied by the protocol required to interpret and reproduce its scores: policy identity, base seed and seed-derivation rule, number of episodes, evaluation-failure count, termination rule, maximum evaluation horizon, and dwell time. For randomized evaluation, the resolved episode seed and sampled generalization parameters are retained for each rollout. Together, these quantities support reproducible audit and reconstruction under a matched simulator, backend, task, and perturbation configuration. F Policy Training Configurations This section discloses the training configurations of the four policies evaluated in Table 2: ACT, DP, ํ 0.5 , and GR00T N1.5. ACT and DP are trained from scratch on Bench2Dex demonstrations;ํ 0.5 and GR00T N1.5 are fine-tuned from their respective public checkpoints. Forํ 0.5 and GR00T N1.5, whose pretrained action heads assume lower-dimensional action spaces than the bimanual-dexterous embodiments require, we retain the pretrained weights and randomly initialize the additional output dimensions to match the target action space. Unless stated otherwise, each policy is trained on a single taskโembodiment setting and consumes the synchronized multi-view RGB and proprioceptive observations recorded in the unified HDF5 episodes. The default hyperparameters reported below are those of the released training launchers (policy/<name>/train.sh); per-task overrides, when applied, are logged with the corresponding checkpoint. ACT. We train ACT from scratch using its conditional variational autoencoder (cVAE) formulation. The encoder and decoder are Transformer networks with a hidden dimension of512and a feed-forward dimension of3,200. Inputs are four camera views (both wrist and both stereo cameras) encoded by a shared ResNet-18 backbone together with the normalized proprioceptive state, whose dimension is sized to each embodimentโs active degrees of freedom. The policy predicts an action chunk of lengthํ pred = 30from a single observation (ํ ํ = 1) and is supervised with anโ 1 reconstruction loss regularized by a KL term of weight10. We optimize with AdamW at a learning rate of1ร 10 โ5 and a batch size of32for up to6,000epochs on a single RTX 4090 GPU. At inference the policy re-plans at every step and fuses overlapping chunk predictions by temporal ensembling. 33 Diffusion Policy (DP).We train DP from scratch using the image-conditioned 1D U-Net variant of Diffusion Policy. Multi-view RGB observations are resized to216ร 288and encoded independently by ResNet-18 visual backbones, whose features are fused into the global conditioning vector of the denoiser. The denoiser is a 1D conditional U-Net with feature widths(256, 512, 1024), a128-dimensional diffusion-step embedding, and kernel size5. At each control cycle the policy conditions on the most recentํ ํ = 3observations, denoises aํ pred = 8-step action trajectory, and executes the firstํ exec = 6actions before re-planning from fresh observations, yielding receding-horizon closed-loop control. Training adopts the DDPM (ํ-prediction) objective over100diffusion timesteps under a squared-cosine noise schedule, and100denoising steps are taken at inference. We optimize with AdamW at a peak learning rate of1ร 10 โ4 , a global batch size of512, 500-step linear warmup followed by cosine decay, and an EMA of the policy weights with decay0.9999. Each task-specific policy is trained for up to 300 epochs on eight H100 GPUs. ํ 0.5 .We fully fine-tune the pretrainedํ 0.5 visionโlanguageโaction model using the openpi implementation, initializing from the publicpi05basecheckpoint. The releasedpi05basedex2benchfullconfiguration trains the non-LoRA PaliGemma-2B visionโlanguage backbone and Gemma-300M action expert jointly. The state and action dimensions are selected from the active degrees of freedom of each embodiment. Compatible pretrained parameters are retained, while state/action projection layers whose shapes differ from the pretrained checkpoint remain randomly initialized. The policy conditions on four RGB views (both stereo and both wrist cameras), normalized proprioception, and the task instruction. It predicts an action chunk ofํ pred = 20 steps with a maximum token length of280; allํ exec = 20actions are executed before the policy re-observes and re-plans. We optimize all trainable parameters with AdamW using a cosine learning-rate schedule with 100warmup steps, a peak learning rate of1ร 10 โ4 , and decay to1ร 10 โ6 over2,000steps. The released launcher uses a global batch size of256,32data-loading workers,fsdpdevices= 1, and a total of2,000 training steps. EMA is disabled. GR00T N1.5. We fine-tune the pretrained GR00T N1.5 generalist policy on the demonstrations of each evaluation task using full-parameter fine-tuning of the backbone, without LoRA adapters. The policy receives four temporally synchronized RGB camera views, the embodiment proprioceptive state, and the task instruction as a per-task language prompt. To accommodate the heterogeneous bimanual-dexterous embodiments, we pad the state vector to a common64-dimensional representation and extend the pretrained action head from its native maximum dimensionality to64dimensions; action channels beyond the pretrained head are newly initialized and optimized jointly with the pretrained parameters, while padded action dimensions are masked from the action objective, preserving a single policy interface across embodiments. GR00T N1.5 uses a flow-matching action head that, from a single observation (ํ ํ = 1), predicts aํ pred = 16-step action chunk; at inference the chunk is generated with4flow-matching denoising steps and executed in full (ํ exec = 16) before the policy re-observes and re-plans. We optimize with AdamW at a peak learning rate of1ร 10 โ4 with5% linear warmup followed by cosine decay and a global batch size of64, training in bfloat16 mixed precision for 20,000 steps on a single H100 GPU. G Generalization Configuration Scene background. Scene background randomization changes the visual surroundings while leaving the tabletop task unchanged. The benchmark includes 60 iTHOR/USD indoor scenes, partitioned into 50 seen and 10 unseen scenes. Each scene has a calibrated yaw and translation offset, and, when this factor is resampled, each episode samples exactly one scene from its designated partition; the nominal background is not substituted. The geometry of the sampled background scene is used only for visual rendering and is excluded from collision and contact simulation. This factor therefore evaluates object grounding under unseen scene backgrounds without introducing room-level physical interactions. 34 Tabletop texture.Tabletop texture randomization changes the visual material of the support surface while preserving its geometry and contact behavior. The texture collection contains 11,824 materials across carpet, fabric, flooring, leather, metal, rust, stone, and wood, with 9,436 seen and 2,388 unseen materials assigned by a deterministic 80/20 split. When this factor is resampled, each episode receives a sampled texture, preventing a fixed tabletop texture from becoming a localization shortcut. Lighting conditions. Lighting randomization modifies the lights embedded in the sampled USD room. The distant light samples intensity in[1000, 1500], each color channel in[0.4, 1.0], pitch in[โ30 โฆ ,โ5 โฆ ], yaw in[โ90 โฆ , 90 โฆ ], and angular size in[0.5 โฆ , 1.0 โฆ ]. The dome light samples intensity in[150, 350], color between[0.8, 0.8, 0.6]and[1, 1, 1], and exposure in[โ0.5, 0.5]. These ranges are shared across splits and probe sensitivity to illumination, shadows, contrast, and exposure. Object pose. Object pose randomization changes task-relevant initial poses within valid bounds. For explicitly positioned objects,ํฅandํฆare independently perturbed by up toยฑ2cm and yaw by up toยฑ10 โฆ . If a sampled pose violates collision or workspace constraints, the perturbation magnitude is progressively reduced; the nominal pose is retained only when no valid perturbed pose is found. Objects without a fixed pose are instead resampled within task-defined zones subject to collision constraints. Task-specific geometric constraints may fix an objectโs initial pose or restrict its perturbation range. Camera pose.Camera pose randomization perturbs external and wrist-mounted camera viewpoints. World- mounted cameras sample translation and look-at-target offsets ofยฑ5m per axis, distance offsets ofยฑ3cm, and rollโpitchโyaw offsets ofยฑ1 โฆ per axis. Wrist cameras sample link-frame translation offsets ofยฑ5m per axis and rotation offsets ofยฑ1 โฆ per axis. Stereo cameras share one perturbation sample to preserve their relative pose. The seen and unseen camera profiles use the same numerical ranges, so this factor evaluates calibration and viewpoint perturbations rather than a disjoint camera-hardware split. Distractor objects.Distractor object randomization adds one to three dynamic distractors per episode. The distractor set contains 68 assets, partitioned into 52 seen and 16 unseen assets; each selected object is scaled by a factor in[0.8, 1.2]. Objects that duplicate a task-relevant object or are semantically confusable with one are excluded before sampling. Placement satisfies collision constraints, a 5 cm table-edge margin, and a central exclusion regionํฅ โ [โ0.25, 0.25]m andํฆ โ [โ0.20, 0.35]m. This procedure evaluates grounding under occlusion, ambiguity, crowding, and incidental contact without deliberately blocking the primary workspace. Table height. Table height randomization samples a vertical offset in[โ0.05, 0.05]m around each taskโs nominal tabletop height (default0.75m). The robot mount remains at its nominal height, while on-table object placement, evaluator height references, and dependent collision checks follow the shifted surface. The resulting change in robot-to-table geometry tests adaptation of reaching, grasping, and contact heights. 35 Table S10. Composition and perturbation strengths of the four Bench2Dex evaluation channels. Entries prefixed by From anchor replay the resolved parameters of the matched anchor episode. Task-level object overrides may narrow or disable pose perturbations. ChannelScene constructionInvariance factors: scene background, tabletop texture, lighting conditions, distractor objects, camera pose Equivariance factors: object pose, table height Evaluation purpose noneMatched anchor; all factors are replayed exactly. From anchor: scene background, tabletop texture, lighting conditions, distractor objects, and camera pose. From anchor: object poses and table height. Establishes the matched baseline without resampling. equionly Matched anchor; only equivariance factors are resampled. From anchor: scene background, tabletop texture, lighting conditions, distractor objects, and camera pose. Resampled: explicit object ํฅโํฆ: ยฑ2 cm; yaw: ยฑ10 โฆ ; objects without fixed pose: task zones; table: ยฑ5 cm. Tests adaptation to task-relevant geometry while controlling visual context. inv onlyMatched anchor; only invariance factors are resampled from the unseen split. Resampled from unseen split: scene background: 10 iTHOR scenes; tabletop texture: 2,388 materials; distant/dome lighting: ranges above; distractor objects: 1โ3 objects from 16 assets at 0.8โ1.2ร scale; world-mounted camera translation/look-at-target offsets:ยฑ5 m, distance: ยฑ3 cm, RPY:ยฑ1 โฆ ; wrist-mounted camera translation: ยฑ5 m, RPY:ยฑ1 โฆ . From anchor: object poses and table height. Tests nuisance robustness under matched task geometry. inv equiIndependent full-scene sample; no anchor is used. Resampled from unseen split: the same perturbations to scene background, tabletop texture, lighting conditions, distractor objects, and camera pose as inv only. Resampled: the same perturbations to object pose and table height as equionly. Tests simultaneous perceptual invariance and geometric adaptation. 36 H Collected Tactile Observations The following figures present representative tactile observations collected across 12 robotโhand embodiments. Each image is shown individually and without cropping. Each figure visualizes the tactile maps of one replayed frame, with one heatmap per tactile site; brighter pixels encode larger quantized penetration depth (values 0โ255, Appendix C.1). Figure S1. Visualization of tactile data collected with the IIWA7+Sharpa embodiment. 37 Figure S2. Visualization of tactile data collected with the Panda+Allegro embodiment. Figure S3. Visualization of tactile data collected with the Panda+Orca embodiment. 38 Figure S4. Visualization of tactile data collected with the RM65+Revo2 embodiment. Figure S5. Visualization of tactile data collected with the UR5+RH56DFX embodiment. 39 Figure S6. Visualization of tactile data collected with the UR5+RH5DG2 embodiment. Figure S7. Visualization of tactile data collected with the UR5+Schunk embodiment. 40 Figure S8. Visualization of tactile data collected with the UR5+Shadow embodiment. Figure S9. Visualization of tactile data collected with the UR5+Wuji embodiment. 41 Figure S10. Visualization of tactile data collected with the xArm7+Ability embodiment. Figure S11. Visualization of tactile data collected with the xArm7+LEAP embodiment. 42 Figure S12. Visualization of tactile data collected with the JAKA ZU7+DexHand021 embodiment. 43