Paper deep dive
ADEPT: Accelerating Dexterity via Pre-Training and Post-Training using Reinforcement Learning
Jayjun Lee, Jessica Yin, Asif Rana, Nicholas Blauch, Sam Mady, Mohak Bhardwaj, Nima Fazeli, Nathan Ratliff, Karl Van Wyk, Ankur Handa
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%
Last extracted: 8/22/2026, 1:45:10 AM
Summary
The paper introduces ADEPT, a reinforcement learning framework that accelerates dexterous manipulation in high-degree-of-freedom robots through pre-training on generic reposing tasks and post-training for specific downstream tasks. It addresses the degradation of pretrained skills during fine-tuning using a stable recipe involving behavior-cloning distillation, critic warm-up, and conservative PPO updates. ADEPT employs a joint-space Geometric Fabric for safety and kinematic dexterity, and distills policies into perceptive students that achieve zero-shot sim-to-real transfer on Kuka-Allegro and Flexiv-Sharpa robots using raw visuo-tactile perception.
Entities (9)
Relation Signals (9)
ADEPT → deployson → Kuka-Allegro
confidence 95% · zero-shot sim-to-real transfer on two embodiments: a 23 DoF Kuka-Allegro
ADEPT → deployson → Flexiv-Sharpa
confidence 95% · and a 29 DoF Flexiv-Sharpa with two RGB cameras and five vision-based tactile sensors
ADEPT → enables → Sim-to-real transfer
confidence 95% · ADEPT enables learning new behaviors... and can solve long-horizon tasks... with dexterity at human-level speed.
ADEPT → uses → Geometric Fabric
confidence 92% · we introduce a joint-space Geometric Fabric that mediates between the RL policy and the robot.
ADEPT → solves → Functional Manipulation Benchmark
confidence 90% · solves the Functional Manipulation Benchmark (FMB) [17] peg insertion task
ADEPT → uses → PPO
confidence 90% · We introduce Accelerating DExterity via Pre-Training (ADEPT), a large-scale reinforcement learning (RL) framework... via PPO
ADEPT → uses → Behavior Cloning
confidence 88% · stable post-training recipe combining behavior-cloning distillation
Flexiv-Sharpa → uses → TacMap
confidence 85% · Following TacMap [29], we simulate each fingertip’s output as a geometry-consistent penetration-depth map
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:We introduce Accelerating Dexterity via Pre-Training (ADEPT), a large-scale reinforcement learning (RL) framework for learning sim-to-real transferable dexterity across high degree-of-freedom (DoF) robot embodiments that can solve long-horizon tasks directly from raw visuo-tactile perception. ADEPT pretrains a dexterous policy on a generic object reposing task, then post-trains downstream policies with this pretrained behavior as a prior. ADEPT enables learning new behaviors that are otherwise difficult to discover from scratch on multi-fingered robots and avoids learning the same set of skills over again for every new downstream task. The pretrained policy zero-shots the reposing phase of downstream tasks, but naïve RL fine-tuning rapidly degrades this capability during transfer. We address this with a stable post-training recipe combining behavior-cloning distillation, critic warm-up, and conservative on-policy updates. To safely exploit the full kinematic dexterity, we introduce a joint-space Geometric Fabric that mediates between the RL policy and the robot. We distill post-trained teachers into perceptive students that zero-shot sim-to-real transfer on two embodiments: a 23 DoF Kuka-Allegro with two RGB cameras, and a 29 DoF Flexiv-Sharpa with two RGB cameras and five vision-based tactile sensors, and can solve long-horizon tasks from challenging initial states with dexterity at human-level speed.
Tags
Links
- Source: https://arxiv.org/abs/2608.19182v1
- Canonical: https://arxiv.org/abs/2608.19182v1
Trouble viewing inline? Open PDF directly →
Full Text
102,293 characters extracted from source content.
Expand or collapse full text
ADEPT: Accelerating Dexterity via Pre-Training and Post-Training using Reinforcement Learning Jayjun Lee Jessica Yin Asif Rana Nicholas Blauch Sam Mady Affiliation: NVIDIA Corporation Affiliation: NVIDIA Corporation Affiliation: NVIDIA Corporation Affiliation: NVIDIA Corporation Affiliation: NVIDIA Corporation Affiliation: Robotics Department, University of Michigan[5pt] adept-dexterity.github.io Mohak Bhardwaj Nima Fazeli Nathan Ratliff Karl Van Wyk Ankur Handa Affiliation: NVIDIA Corporation Affiliation: NVIDIA Corporation Affiliation: NVIDIA Corporation Affiliation: NVIDIA Corporation Affiliation: Robotics Department, University of Michigan[5pt] adept-dexterity.github.io Abstract We introduce Accelerating DExterity via Pre-Training (ADEPT), a large-scale reinforcement learning (RL) framework for learning sim-to-real transferable dexterity across high degree-of-freedom (DoF) robot embodiments that can solve long-horizon tasks directly from raw visuo-tactile perception. ADEPT pretrains a dexterous policy on a generic object reposing task, then post-trains downstream policies with this pretrained behavior as a prior. ADEPT enables learning new behaviors that are otherwise difficult to discover from scratch on multi-fingered robots and avoids learning the same set of skills over again for every new downstream task. The pretrained policy zero-shots the reposing phase of downstream tasks, but naïve RL fine-tuning rapidly degrades this capability during transfer. We address this with a stable post-training recipe combining behavior-cloning distillation, critic warm-up, and conservative on-policy updates. To safely exploit the full kinematic dexterity, we introduce a joint-space Geometric Fabric that mediates between the RL policy and the robot. We distill post-trained teachers into perceptive students that zero-shot sim-to-real transfer on two embodiments: a 23 DoF Kuka-Allegro with two RGB cameras, and a 29 DoF Flexiv-Sharpa with two RGB cameras and five vision-based tactile sensors, and can solve long-horizon tasks from challenging initial states with dexterity at human-level speed. Figure 1: ADEPT enables sim-to-real RL of high DoF arm-hand robots to reach, grasp, lift, reorient, transport, align, and insert. We demonstrate on a 23 DoF Kuka-Allegro from two RGB cameras and on a 29 DoF Flexiv-Sharpa from two RGB cameras and five fingertip tactile sensors. Keywords: Multisensory Dexterity, Sim-to-Real Reinforcement Learning 1 Introduction Dexterous manipulation with high degrees-of-freedom robotic systems, such as arms equipped with anthropomorphic multi-fingered hands, remains a major challenge for robot learning [1, 9, 16, 4]. Multi-fingered arm-hand systems combine high-dimensional state and action spaces with contact-rich interactions, making useful behaviors hard to discover from sparse task-specific rewards. Even with massively parallel GPU simulation [18], RL policies trained from scratch for one task rarely transfer to another; each new task starts over, rediscovering the same low-level skills like reaching, grasping, lifting, and reorientation before any task-specific behavior can emerge. This motivates a different paradigm. Rather than training an RL specialist from scratch for each new dexterous task, we propose first acquiring a general-purpose dexterous foundation and then adapting it for specific downstream tasks [14, 19]. The policy’s initialization in parameter space is as consequential as the downstream objective it must adapt to. A generic pretraining task that exercises the underlying motor primitives lands the network in a region from which task-specific behaviors build on, rather than overwrite, the pretrained skills [13]. We introduce Accelerating DExterity via Pre-Training (ADEPT), a framework that pretrains foundational arm-hand-object dexterity on a generic object reposing task [20] and post-trains it for downstream contact-rich tasks. We find that this pretrained policy zero-shots the reposing segment of downstream tasks despite never having seen them during training, but transferring the policy and fast adaptation to a contact-rich downstream task is non-trivial. Naïve RL fine-tuning often collapses the pretrained behaviors before they can be specialized to the new task [27]. We trace this failure to mismatched observation spaces, misaligned value estimates, and excessive policy drift during early PPO [26] updates, and address it with a structured post-training procedure of actor distillation, critic warm-up, and conservative on-policy updates, respectively. We finally distill post-trained teachers into deployable end-to-end perceptive students through a two-stage distillation curriculum and deploy them zero-shot on two arm-hand embodiments. On a 23 DoF Kuka iiwa7 + Allegro system, an RGB student solves the Functional Manipulation Benchmark (FMB) [17] peg insertion task end-to-end as a single learned policy on two peg geometries that span the benchmark’s difficulty range: a symmetric star peg and the asymmetric square-and-round peg, the most challenging geometry in FMB, and additionally solves dish-rack placement. On a 29 DoF Flexiv-Sharpa system, a visuo-tactile student solves FMB peg insertion from two RGB cameras and five fingertip tactile sensors. FMB was originally set up for parallel-jaw grippers, which require external fixtures to reorient and multiple stages of interaction with the object, both of which our approach eschews by using multi-fingered hands while discovering in-hand dexterous manipulations not easily possible with human demonstrations. To summarize, our contributions are: • ADEPT, a framework for pretraining foundational arm–hand–object dexterity on a generic reposing task using RL and post-training it for contact-rich downstream tasks that accelerates and makes full DoF dexterous state-based RL tractable. • A novel full joint configuration space (Cspace) geometric fabric [30] that gives the RL policy access to the full kinematic dexterity of high-DoF arm–hand systems while enforcing hardware safety (preventing collisions and hitting joint limits), in contrast to prior fabric-guided dexterous policies [16, 28] that restrict the hand to a low-dimensional PCA grasp subspace. • A structured RL post-training recipe and distillation curriculum that together enable zero-shot sim-to-real on contact-rich, long-horizon dexterous insertion tasks across multiple arm-hand embodiments. Our results represent the first demonstration of challenging pick-reorient-insert with robotic arm-hand systems via sim-to-real RL from raw perception without demonstrations or pose trackers, on a vision-based Kuka-Allegro and a visuo-tactile Flexiv-Sharpa. Each task and embodiment is trained independently, using the same pre-trained policies specific to that embodiment. • A 2×2×–14×14× execution-time speedup over the FMB [17] parallel-jaw pipeline: our multi-fingered policies solve each task in 55–1010 s per trial, compared to 2020–7070 s per trial for the FMB parallel-jaw pipeline that relies on external fixtures and multi-stage regrasp decomposition. 2 Related Work Sim-to-Real Dexterous Manipulation. [2, 1] established large-scale PPO with domain randomization for sim-to-real dexterity, followed by in-hand reorientation works [9, 5, 4, 21, 22]. ManipGen [7] composes local sim-only policies with motion planners for long-horizon real-world tasks, and OmniReset [31] diversifies reset-state distributions to elicit emergent dexterity from PPO, a form of generality that we instead achieve at the policy level through reposing pretraining. SimToolReal [12] similarly trains a task-agnostic object-centric RL policy on procedurally generated primitives for zero-shot real-world tool use, but its deployed state-based policy is explicitly conditioned on estimated object and goal poses. ADEPT instead deploys perception-based students that consume raw RGB and fingertip tactile directly, without a pose estimator in the loop. DextrAH-G and DextrAH-RGB [16, 28] use geometric fabrics as a low-level controller for sim-to-real grasping. DemoStart [3] learns grasp–reorient–insert behaviors on a three-fingered hand from pixels but requires demonstrations to seed learning. To our knowledge, prior work has not demonstrated zero-shot sim-to-real RL for long-horizon pick–reorient–insert on a high-DoF arm–hand system directly from raw visual or visuo-tactile perception, without demonstrations or a pose estimator. Reinforcement Learning Pretraining and Transfer. Prior work studies pretraining and fine-tuning for RL [14], including behavior priors and regularization methods that mitigate distribution shift or forgetting during online adaptation [33, 32, 19, 27, 13]. ADEPT operates in this regime but is tailored to high-DoF dexterous policies trained from scratch in simulation: rather than regularizing with an explicit KL or EWC penalty, ADEPT initializes the downstream actor through behavior distillation, warms up the critic, and uses a substantially reduced actor learning rate during PPO post-training. Contemporaneously with our work, Play2Perfect [15] pretrains a goal-conditioned RL policy through task-agnostic dexterous play and fine-tunes it on precise contact-rich assembly, but, like SimToolReal [12], takes object pose as an input observation rather than raw perception and deploys state-based teachers in the real world. Geometric Fabrics for Policy Learning. Geometric fabrics [30] provide a smooth, second-order action prior that RL policies command at the acceleration level, yielding reactive motions with provable stability guarantees and built-in collision and joint-limit avoidance, extending a line of work on geometrically consistent reactive control [23, 6, 24]. DextrAH-G [16] and DextrAH-RGB [28] use geometric fabrics for sim-to-real grasping but restrict hand control to a 5D PCA subspace of retargeted human grasps plus a 6D palm pose target, limiting the finger coordination needed for contact-rich manipulation. ADEPT instead drives geometric fabrics in the full joint configuration space (Cspace fabrics), exposing the full arm–hand kinematic space to the policy at the cost of a substantially harder learning problem. Tactile Sensing for Sim-to-Real Dexterous Manipulation. Prior sim-to-real dexterous policies have largely relied on vision alone, since simulated contact signals cross the sim-to-real gap far less cleanly than images. Recent work bridges this gap by grounding both domains in a shared representation. HydroShear [8] simulates hydroelastic shear for vision-based tactile sensors, enabling zero-shot sim-to-real manipulation with parallel-jaw grippers. TacMap [29] instead represents tactile observations as geometry-consistent penetration-depth maps shared between simulation and hardware. SaTA [10] spatially anchors per-finger tactile features using FiLM conditioning on fingertip positions. ADEPT combines TacMap-derived penetration-depth and binary contact maps with SaTA-style FiLM anchoring on fingertip positions to provide explicit, spatially disambiguated per-finger contact information for sim-to-real dexterous manipulation. 3 Methodology We propose a framework for pretraining foundational dexterity on a generic task objective and post-training for contact-rich downstream tasks as outlined in Fig. 2. Our approach consists of three stages: (1) pretraining a dexterous policy in simulation on a generic object manipulation task, (2) transferring the pretrained behavior through distillation and adapting through massively parallel on-policy RL post-training while preventing catastrophic forgetting through stable policy updates, and (3) teacher-student distillation using a task specialist teacher distilled into a stereo RGB vision-based student that can be zero-shot deployed in the real-world. Figure 2: ADEPT Overview. (1) Pre-train (πpre,Vpre)( _pre,V_pre) via PPO on a generic reposing task. (2) Post-train into (πpost,Vpost)( _post,V_post) for the downstream contact-rich task via BC distillation, frozen actor critic warm-up, and conservative PPO. (3) Distill πpost _post into a stereo RGB student πstudent _student. (4) Deploy πstudent _student zero-shot on the real robot. 3.1 Problem Formulation We cast dexterous manipulation as a discrete-time Markov decision process (MDP) ℳ=(,,p,r,γ)M=(S,A,p,r,γ) with continuous state space S, action space =[−1,1]nqA=[-1,1]^n_q (nq=23n_q=23), transition dynamics p, scalar reward r, and discount factor γ. We learn a stochastic policy πθ(at∣ot) _θ(a_t o_t) that maximizes the expected discounted return J(θ)=τ∼πθ[∑t=0Tγtr(st,at)]J(θ)=E_τ _θ [ _t=0^Tγ^tr(s_t,a_t) ] via PPO [26] with an asymmetric actor-critic. ADEPT operates over two MDPs ℳpreM_pre and ℳpostM_post for the pretraining (Sec. 3.2) and post-training (Sec. 3.3) tasks respectively. They share the action space A but differ in observation space, dynamics, and reward; in particular, oposto^post extends opreo^pre with task-specific signals: receptacle pose recp_rec and object–receptacle contact forces orf_or. We first solve ℳpreM_pre with PPO to obtain a pretrained actor-critic (πpre,Vpre)( _pre,V_pre), then bootstrap from this pretrained actor to learn a downstream actor–critic (πpost,Vpost)( _post,V_post) on ℳpostM_post via the structured post-training of Sec. 3.3. 3.2 Pre-Training Foundational Dexterity We pretrain a dexterous teacher actor-critic with PPO on a generic object reposing task designed to instill broad manipulation capabilities. At each episode, one of 16 primitive shapes (cylinders, cuboids, spheres, cones) at a randomized scale is spawned on the table, and the policy must execute a full sequence of reaching, grasping, lifting, in-hand reorientation, transporting, and reposing the object to a sampled target goal pose. Following [1, 9], we adopt ADR as an online curriculum that advances task goal and increases environmental complexity as the agent’s success rate improves, and additionally use Population-Based Training (PBT) [11, 20] (Appx. C) to search over PPO hyperparameters, and gravity is ADR-annealed from 00 to −9.81-9.81 m/s2. This setup encourages the emergence of reusable dexterous skills (precise grasps, coordinated finger motion, and in-hand reorientation across varied object geometries and scales), yielding foundational dexterity that solves the reposing segment of downstream tasks in a zero-shot manner (Tab. 2). Refer to Appx. A.1 for the full task specification, observation space, and reward function. Note that we use object point clouds to represent objects, which has shown better zero-shot generalization for downstream tasks. 3.3 Post-training Dexterous RL Specialists on Downstream Tasks Challenges in Transferring Pretrained Dexterity. Directly fine-tuning the pretrained policy on a downstream task using standard reinforcement learning leads to rapid degradation of its behavior as shown in Sec. 4.1 by the inset of Fig. 3 where it loses the ability to zero-shot the reposing segment of the task immediately after some policy updates (0% success). We observe that this degradation occurs even when the pretrained policy exhibits strong zero-shot performance, indicating that the failure arises during training rather than from insufficient pretraining. We attribute this to a number of mismatches under the transfer learning setup: (1) change in task reward function, (2) extra observations that are only available in the downstream task (e.g. contact forces between object-receptacle and pose of the receptacle), (3) poor value or advantage estimates, and thus (4) large policy updates. Structured RL Adaptation. We frame downstream learning as layering new behaviors on top of the pretrained dexterous prior rather than relearning manipulation from scratch: each iteration of on-policy PPO is restricted to a local region of the pretrained policy space, so the policy retains core reaching, grasping, lifting, and reorientation skills while incorporating downstream task-specific observations recp_rec and orf_or and reward. Instead, ADEPT stabilizes transfer through actor distillation, critic calibration, and conservative policy updates while still allowing downstream behavioral adaptation. Concretely, we transfer the pretrained actor–critic (πpre,Vpre)( _pre,V_pre) into the downstream task in three steps. (1) BC actor distillation. We distill πpre _pre into a new downstream actor πpost _post with the downstream observation space using supervised imitation for 40k iterations. (2) Critic warm-up. We then freeze πpost _post and train a fresh downstream critic VpostV_post on rollouts collected by the frozen πpost _post under the downstream reward, aligning value estimates with the new task over 20 PPO iterations (∼1 1M env steps per GPU with 4096 envs) before any policy updates are applied. (3) Conservative PPO. We finally unfreeze πpost _post and jointly update (πpost,Vpost)( _post,V_post) with PPO using conservative updates (actor LR 1e-3 → 1e-5 with linear decay, PPO clip ϵε: 0.2→0.050.2→ 0.05, critic LR fixed at 5e-5). BC distillation (1) runs against the single ADR 20 goal pose on the goal path illustrated in Fig. 7 by grey spheres; critic warm-up (2) and conservative PPO (3) then train directly against the final insertion goal at ADR 50 while the remaining ADR randomizations anneal from level 20; training from scratch must follow the same curriculum from ADR 0. Controlled ablations (Tab. 3) show that the reduced actor learning rate is what prevents collapse, while BC and critic warm-up improve success, stability, and adaptation speed; the tightened clip has little effect. Refer to Alg. 1 for the full algorithm and Tab. 5 in Appx. A.3 for the training hyperparameters. 3.4 Teacher-Student Distillation Setup. After RL post-training yields a state-based teacher for a contact-rich downstream task, we distill it into a deployable student that consumes only proprioception, fabric state, two RGB images, and, on the Flexiv-Sharpa embodiment, five per-finger tactile maps as outlined in Tab. 6 in Appx. A.4 and can be deployed zero-shot on the real robot. The student is trained with DAgger [25]: the teacher provides per-step action targets (mean and variance) over rollouts collected from the student in the same task environment, and the student uses an initially frozen pretrained ResNet backbone with a DextrAH-RGB-style [28] unfreeze schedule. Auxiliary loss. The action-cloning loss alone does not reliably teach the student to perceive the peg’s orientation, which is the dominant failure mode on our insertion task: the peg must be inserted at a specific orientation, so small errors in predicted yaw or tilt translate directly into missed insertions. Following DextrAH-RGB [28], which uses object position prediction as the auxiliary loss, we add an 8-keypoint object pose prediction head on top of the shared stereo image features and supervise it against the ground-truth peg pose available in simulation. The full distillation objective is the sum of the BC and auxiliary losses, ℒ=ℒBC+ℒauxL=L_BC+L_aux, with more details in Appx. G. Two-stage student curriculum. Mirroring the pretraining/post-training structure of our RL pipeline, we introduce a vision-side student pretraining stage that isolates the perceptual prerequisites of the downstream task (the peg detection and 8-keypoint pose prediction that ℒauxL_aux supervises) from the policy-learning objective driven by ℒBCL_BC. (1) Student pretraining via the reposing teacher. We re-task the pretrained reorientation teacher (Sec. 3.2) to perform a perception-heavy surrogate of the downstream task (lifting and reorienting the peg to an upright pose at a fixed receptacle location) and distill it into the stereo RGB student. The student must learn from RGB alone to detect the peg, predict its 8-keypoint pose as an auxiliary prediction, track it through the grasp and lift, and reproduce the teacher’s reorientation behavior. (2) Vision distillation from the downstream teacher. We then initialize the student in the downstream distillation from the Stage-1 checkpoint and continue training against the post-trained downstream teacher on the insertion task. This two-stage curriculum keeps a single dominant learning objective at each stage: ℒauxL_aux shapes the visual encoder in Stage 1 while ℒBCL_BC then drives contact-rich policy refinement in Stage 2 on top of an already-competent encoder, removing the conflict between perception and policy learning that arises when the two are trained jointly from scratch. Tactile perception for visuo-tactile students. For the Flexiv-Sharpa robot, we extend the student’s observation space with per-finger tactile images from all five fingertip vision-based tactile sensors. Following TacMap [29], we simulate each fingertip’s output as a geometry-consistent penetration-depth map that shares its representation with the calibrated real-sensor output, so the student consumes tactile signals without additional sim-to-real bridging. From each depth map we derive a binary contact map by thresholding, encode both channels with a per-finger CNN shared across fingers, and spatially anchor the resulting features via SaTA-style [10] FiLM conditioning on the fingertip position obtained via forward kinematics. The per-finger tactile embeddings are then fused with the stereo RGB and proprioception latents, giving the policy a spatially grounded view of hand-object contact. Full architecture and dimensions are deferred to Appx. H. Randomization. Both distillation stages use aggressive domain randomization on physics, object perturbations via random wrenches, visual conditions (lighting, background, camera intrinsics and pose), and sensor noise on proprioception and contact signals. Refer to Appx. I for more details. 3.5 Joint Configuration Space Geometric Fabrics We place a geometric fabric [30] between the policy and the robot to enforce hardware constraints (joint limits, self- and environmental collision), provide a smooth second-order action prior that shapes RL exploration, and guarantee the same low-level controller in simulation and on the real robot. A fabric is an autonomous second-order system on configuration space, f(f,˙f)¨f+f(f,˙f)+π()=M_f(q_f, q_f)\, q_f+f_f(q_f, q_f)+f_π(a)=0, where ff_f collects autonomous geometric and dissipative terms (collision and joint-limit repulsion, Cspace damping, speed control) and π()f_π(a) is the policy-driven forcing term. To support behaviors such as in-hand reposing, finger gaiting, and contact-rich manipulation, we drive the fabric in the full joint configuration space: the policy outputs t∈[−1,1]nqa_t∈[-1,1]^n_q (nq=23n_q=23 for Kuka-Allegro), interpreted as per-joint relative deltas that map to a Cspace target consumed by πf_π. This exposes the full 2323-DoF kinematic dexterity of the arm–hand system to the policy while retaining the fabric’s safety and stability guarantees [30]. The same fabric instance runs in simulation and on the real robot, minimizing the controller gap from the sim-to-real problem. Full fabric components, the relative-target mapping, and the integration loop are deferred to Appx. B. 4 Experiments 4.1 Simulation Experiments How well does the pretrained teacher generalize to unseen objects? We evaluate the pretrained teacher on the reposing task across three object categories in Tab. 1: the 16 in-distribution primitive shapes used during pretraining, the FMB pegs (star and square/round) from downstream insertion task, and VisDex [4] objects. Despite training only on primitives, the Kuka-Allegro teacher matches or slightly exceeds its in-distribution success rate on both out-of-distribution (OOD) sets, and the Flexiv-Sharpa teacher stays close to the in-distribution success rate rather than failing to generalize. How far can the pretrained teacher zero-shot the downstream task? We evaluate the teacher on FMB peg insertion across ADR levels with no post-training in Tab. 2. Goal poses are laid out along an L-shaped path indexed by ADR level, as visualized by the grey spheres in Fig. 7 in Appx. A.2: a horizontal lift-and-transport segment from above the peg spawn to above the board hole (ADR 0 to 25), then a vertical insertion segment descending into the hole (ADR 25 to 50). Zero-shot success stays above 50%50\% through ADR 3535 and reaches 0%0\% at the actual insertion goal (ADR 5050) as it starts making contact with the receptacle, which is OOD. We therefore begin all downstream post-training at ADR 20: BC distillation clones the teacher against the ADR 20 goal pose, and from critic warm-up onward the goal pose is set directly to the final insertion goal at ADR 50, the endpoint of the path, while the remaining ADR randomizations anneal from level 20. For FMB peg insertion the pretrained policy zero-shots the reposing portion reliably at this level, so only the contact-rich insertion segment remains to be learned. For dish placement the reposing prior transfers only up to reaching, not grasping, since the 16 primitive pretraining objects include nothing close to the plate’s large and flat geometry. We later show that post-training nevertheless learns the full task. Table 1: Episodic success rates (SR) of pre-trained reposing teacher on unseen objects. Mean ± std across seeds at the final ADR level. Metric Primitive FMB Peg VisDex [4] (Fixed Goal) Number of objects 16 2 152 Kuka-Allegro SR ↑ 0.73±0.0030.73± 0.003 0.76±0.0030.76± 0.003 0.77±0.0110.77± 0.011 Flexiv-Sharpa SR ↑ 0.64±0.0070.64± 0.007 0.58±0.0110.58± 0.011 0.61±0.0150.61± 0.015 Table 2: Zero-shot success rate of the pretrained reposing teacher on FMB peg reposing across ADR. Zero-shot Success Rate (%) at ADR Level across 1024 Episodes Task (Embodiment) 0 5 10 15 20 25 30 35 40 45 50 FMB Peg (Kuka-Allegro) 98.698.6 94.194.1 84.784.7 78.678.6 71.971.9 67.967.9 58.758.7 52.552.5 39.639.6 18.018.0 0.000.00 FMB Peg (Flexiv-Sharpa) 94.594.5 87.187.1 80.380.3 76.876.8 74.974.9 70.470.4 66.866.8 59.259.2 52.352.3 22.922.9 0.700.70 Figure 3: Training curves on FMB peg insertion comparing (1) training from scratch, (2) direct PPO finetuning of the pretrained reposing teacher, and (3) ADEPT post-training. Blue inset: success-rate collapse for (2) during transfer. How does ADEPT compare to training from scratch? Fig. 3 plots ADR level over environment steps for (1) training from scratch on the downstream task, (2) direct PPO finetuning of the pretrained reposing teacher with extra observations, and (3) ADEPT post-training. Fig. 8 in Appx. E shows the same plot with wall-clock time. ADEPT trains a downstream-task teacher in 3B environment steps on top of the 8B-step reposing pretraining, for a total of 11B steps. Training from scratch is highly seed-sensitive, with most seeds plateauing well below ADR 6 as shown in Fig. 9 in Appx. E and only a few reaching it after about 9B steps. More importantly, the 8B pretraining cost is amortized across every downstream task, so the marginal cost per new task reduces to the 3B post-training. Why does naïve RL finetuning fail? The blue inset of Fig. 3 zooms in on the success rate at ADR level 20, where (2) and (3) both begin from the same pretrained teacher: naïve PPO finetuning drives the success rate to zero within a small number of updates, while ADEPT post-training continues to improve. The pretrained critic VpreV_pre is calibrated to the reposing reward, so under the new insertion reward its value estimates are off and the advantages it feeds to PPO are unreliable. The resulting policy gradients push the actor away from its pretrained distribution faster than the critic can recalibrate; each subsequent rollout is collected from a worse policy operating on a worse critic, and within a few iterations the pretrained behavior is destroyed. ADEPT’s post-training addresses this directly: BC actor distillation transfers the pretrained teacher actor’s knowledge into the downstream actor with the new observation space oposto^post, critic warm-up recalibrates VpostV_post under the new reward with the actor frozen so subsequent policy updates see reliable advantages, and conservative PPO bounds the on-policy update to prevent large drifts from the pretrained policy manifold. Which post-training components matter? Controlled ablations in Tab. 3 decompose the recipe over five independent (non-PBT) seeds per variant, all initialized from the same pretrained checkpoint. The reduced actor learning rate is the necessary component: every 10−310^-3 variant collapses at ADR 20 with zero success, even with BC and critic warm-up in place. A KL penalty DKL(πpost|πpre)D_KL( _post| _pre), which regularizes the post-training policy does not rescue it, and at a learning rate of 10−510^-5, adding the penalty with β=1β=1 still yields 0%0\% across all five seeds. Given the low learning rate, critic warm-up is worth +17.6+17.6 SR (row c) and BC nearly halves adaptation time (19.919.9 vs. 35.235.2 h, row b); removing both stalls every seed below ADR 40. Without BC, the ten new observation dimensions must be injected into the pretrained actor, and we test both constructions: row (g) remaps the actor input layer (nine obsolete inputs removed, ten zero-initialized inputs added), while row (g†) keeps all 391 pretraining inputs and appends the ten. Both stall, so the gap is not an artifact of the remapping. The tightened clip has little effect: loosening it to 0.200.20 (row f) matches or exceeds the full recipe on every column. All policies deployed in this paper use the tight clip, and the ablation makes the loose clip a sound default for future post-training runs. Table 3: Post-training component ablations on KUKA-Allegro FMB peg insertion; 55 independent non-PBT seeds per row, all from one pretrained checkpoint. ADR50 counts seeds reaching the final ADR level; time is averaged over seeds that reached it. Train SR is the instantaneous training success rate over parallel environments. Variant BC WU LR Clip ADR50 Train SR (%) Time to ADR 50 (a) ADEPT (full) ✓ ✓ 10−510^-5 .05.05 5/55/5 46.0±1.146.0± 1.1 19.9±1.019.9± 1.0 h (b) No BC ✗ ✓ 10−510^-5 .05.05 4/54/5 38.5±1.238.5± 1.2 35.2±4.935.2± 4.9 h (c) No warm-up ✓ ✗ 10−510^-5 .05.05 4/54/5 28.4±3.728.4± 3.7 20.2±1.520.2± 1.5 h (d) Standard PPO ✓ ✓ 10−310^-3 .20.20 0/50/5 0.0±0.00.0± 0.0 Collapse (ADR 20) (e) High LR only ✓ ✓ 10−310^-3 .05.05 0/50/5 0.0±0.00.0± 0.0 Collapse (ADR 20) (f) Loose clip ✓ ✓ 10−510^-5 .20.20 5/55/5 46.8±2.046.8± 2.0 17.6±1.817.6± 1.8 h (g) No BC/WU (remapped) ✗ ✗ 10−510^-5 .05.05 0/50/5 26.6±7.926.6± 7.9 Stall (ADR 29–39) (g†) No BC/WU (append-only) ✗ ✗ 10−510^-5 .05.05 0/50/5 27.0±10.727.0± 10.7 Stall (ADR 43) (h) Direct FT ✗ ✗ 10−310^-3 .20.20 0/50/5 0.0±0.00.0± 0.0 Collapse (ADR 20) (i) Direct FT + KL ✗ ✗ 10−310^-3 .20.20 0/50/5 0.0±0.00.0± 0.0 Collapse (ADR 20) Emergent natural grasps from ADEPT and refinement from post-training. ADEPT offers more than accelerated dexterous policy learning and a stable post-training framework: it has the emergent property of synthesizing natural grasps. The insertion reward does not constrain how the peg is grasped; both natural and unnatural grasps can satisfy it. From a random initialization, PPO converges to whichever policy maximizes the reward, often synthesizing an unnatural grasp as in Fig. 12(c) and performance varies significantly across seeds. What truly produces the natural grasps is reposing pretraining, which initializes the policy in a region of policy parameter space where natural grasps emerge. Post-training then applies local policy updates from this initialization against the same reward; rather than discovering new grasp modes, it refines the pretrained natural grasps to specialize for the downstream task. Task-misaligned grasps that the reposing teacher occasionally produced in Fig. 12(b) (e.g., grasping from the bottom of the peg at ADR level 20) disappear under post-training refinement as guided by the downstream hand-to-object and object-to-goal alignment reward. Refer to Appx. J for representative ADEPT, pretrained teacher, and from-scratch grasps. The plate in the dish task is larger and flatter than any object seen during pretraining, and no pretrained grasp works on it. Starting from reaching that transfers zero-shot and the policy’s natural grasping attempts, post-training learns the entire new task with the downstream reward. Can we learn novel behaviors through post-training? If post-training could only refine behaviors already present in the pretrained policy, its usefulness would be limited by pre-training coverage. The dish task shows that post-training can go beyond straightforward refinement. Reaching transfers zero-shot: the pretrained policy reliably approaches the plate. Grasping, however, does not. Unlike the FMB pegs, which remain within the pretrained policy’s grasp repertoire (Tab. 1 and Fig. 12), the plate is geometrically far from the 16 simple primitives (Fig. 6) used during pre-training, and none of the pretrained grasps succeed on it zero-shot. Post-training therefore begins without a competent grasping behavior to preserve or refine. What pre-training provides instead is a useful starting distribution: the policy already reaches the plate and produces plausible, though unsuccessful, grasp attempts near meaningful contact configurations. From this initialization, downstream RL learns successful plate grasping and, for upside-down initializations, a flip-and-regrasp strategy not successfully exhibited by the pretrained policy. This contrasts with FMB, where post-training primarily selects and refines grasp modes already available in the pretrained policy (Fig. 12, Appx. J). Thus, pre-training need not already solve the downstream behavior; it is sufficient for the prior to place the policy in a region from which downstream RL can discover a successful solution. For tasks requiring qualitatively different manipulation skills with little overlap with the pretrained prior, additional pre-training coverage may be necessary. 4.2 Real World Experiments Figure 4: Real-world setup. (A) The 23 DoF Kuka-Allegro and 29 DoF Flexiv-Sharpa platforms, each with two RealSense RGB cameras (left and center). (B) FMB pegs and boards, plates, and dish rack. (C) Diverse initial states and lighting used in real-world testing. Hardware Setup. We deploy on two arm-hand platforms: a 7 DoF Kuka iiwa7 with a 16 DoF Allegro four-fingered hand (nq=23n_q=23), and a 7 DoF Flexiv Rizon with a 22 DoF Sharpa five-fingered hand (nq=29n_q=29), each rigidly mounted to a workbench and observed by two calibrated Intel RealSense RGB cameras in the same left-and-center layout (Fig. 4). Each Sharpa fingertip additionally carries a vision-based tactile sensor whose deformation is exposed to the policy as per-finger TacMap depth maps, the same representation rendered in simulation during training. Student policies run on a separate workstation and zero-shot transfer from simulation by sharing the same geometric-fabric action interface (Sec. 3.5), and, on Flexiv-Sharpa, the same tactile interface, in both simulation and on the real robot. We provide the full deployment topology in Appx. F, sample real-world camera views in Fig. 10, and the visuo-tactile student architecture in Appx. H. Task Setup. We evaluate on two FMB [17] pegs in Fig. 4 that span the benchmark’s geometric difficulty: a symmetric star peg, whose rotational and top-down symmetry admits multiple valid insertion orientations, and the asymmetric square-and-round peg, the most challenging geometry in FMB, whose square and round legs admit only a single valid insertion orientation. We initialize pegs flat on the table over 30 cm × 25 cm in spawn range as illustrated in Fig. 7 and [−π,π]rad[-π,π]\ rad in yaw and count a trial successful when the peg is fully inserted. On Kuka-Allegro we additionally evaluate dish-rack placement (Fig. 4): the policy must grasp a plate, reorient it upright, and seat it into a slot of a wooden dish rack, a flip-and-regrasp pattern that never appears in reposing pre-training. Plates are initialized flat on the table, either upright or upside down, over a spawn range similar to the FMB pegs, and a trial counts as successful when the plate is fully seated in a rack slot. In the per-stage breakdown of Tab. 4, the flip-and-regrasp of upside-down plates is counted as part of the Grasp stage. Results. A single Kuka-Allegro reposing checkpoint post-trains to three downstream tasks, FMB star insertion, FMB square/round insertion, and dish-rack placement, with one post-trained teacher per task and no task-specific pre-training. Zero-shot in the real world, the vision student reaches 5/105/10 (star) and 3/103/10 (square/round) on FMB and 6/106/10 on dish placement; the visuo-tactile Flexiv-Sharpa student reaches 8/108/10 on FMB square/round, the only peg trained on that embodiment (Tab. 4). Post-trained teachers evaluate at 85.0%85.0\% (Kuka-Allegro, aggregated over both pegs) and 89.2%89.2\% (Flexiv-Sharpa, square/round) in simulation, and the Kuka-Allegro vision student evaluates at 46.7%46.7\%/65.2%65.2\% (square/round, star). All numbers here are episodic success rates over 10241024 episodes at ADR 5050; training curves elsewhere report the instantaneous fraction of environments in the success state, which resets after sustained success and is not comparable to episodic evaluation. Star vs. square/round. The star peg’s ridges allow fingers to wrap between them for stable grasps, and its symmetry admits multiple valid insertion orientations. The asymmetric peg’s round leg yields unstable point contacts, producing the largest per-stage drops at lifting and reorienting (Tab. 4) and a lower overall success rate. Vision vs. visuo-tactile. In real deployment, tactile sensing is decisive (8/108/10 vs. 3/103/10 overall on Flexiv-Sharpa). The failure mode is not grasp execution but grasp confidence: the vision-only student cannot tell whether a grasp has succeeded, so even after a good grasp it often reopens the hand, drops the object, and loops between grasping and regrasping, and these failures cascade through lifting and reorientation (Tab. 4). With fingertip tactile sensing the policy is certain of its contacts: it grasps and lifts successfully in every trial and carries that reliability through reorientation, alignment, and insertion. Distillation curriculum. The single-stage baseline fails to transfer entirely (0/100/10 on both pegs), while the two-stage curriculum outperforms it by roughly 10%10\% in instantaneous success rates across parallel environments throughout training, confirming that Stage-1 perception pretraining is critical for sim-to-real transfer. Human-level speed. Our policies execute the entire manipulation sequence as a single continuous behavior at human-comparable timescales (55–1010 s per trial), reorienting objects in-hand with a single hand. FMB [17] human demonstrations with a parallel-jaw gripper instead decompose the task into grasp, place-on-fixture, regrasp, rotate, and insert primitives, rely on external fixtures, and require 2020–7070 s per trial, a 2×2×–14×14× execution-time difference. Table 4: Per-stage success rates on real-world long-horizon manipulation, trained with ADEPT. Each stage is cumulative: success at stage k requires success at all earlier stages. Modality Robot Task Reach Grasp Lifting Reorient Align Insert (SR) Vision Kuka-Allegro FMB Star 10/10 9/10 8/10 8/10 7/10 5/10 Vision Kuka-Allegro FMB Square/Round 10/10 8/10 6/10 4/10 3/10 3/10 Vision Flexiv-Sharpa FMB Square/Round 10/10 7/10 5/10 3/10 3/10 3/10 Visuo-Tactile Flexiv-Sharpa FMB Square/Round 10/10 10/10 10/10 9/10 8/10 8/10 Vision Kuka-Allegro Dish 10/10 10/10 8/10 7/10 6/10 6/10 5 Limitations Although the real-world policies exhibit behaviors closely matching those observed in simulation, perception remains the primary bottleneck for distillation. Failures frequently coincide with incorrect estimates of the asymmetric peg’s orientation under occlusion, while grasp instability can arise from the small contact patch between the rounded Allegro fingertips and the rounded side of the peg under fast arm motion. Fingertip tactile sensing on Flexiv-Sharpa mitigates contact ambiguity, but robust object-centric perception under occlusion remains open. A wrist-mounted camera and tactile sensing on additional platforms may help address these limitations. 6 Conclusion and Discussion We presented ADEPT, a pre-training and post-training framework for learning dexterous arm-hand manipulation with reinforcement learning. A generic reposing task provides a reusable initialization for downstream manipulation, while structured post-training preserves useful pretrained behaviors as the policy adapts and aligns to new contact-rich task objectives. Across Kuka-Allegro and Flexiv-Sharpa, the same approach supports long horizon behaviors including grasping, reorientation, transport, and insertion, and transfers to real robots through vision and visuo-tactile student policies. Our results suggest that reusable dexterous pre-training can reduce how much must be rediscovered for each new manipulation problem, while still leaving room for downstream RL to adapt and acquire task-specific behavior. An important next step is to broaden the pre-training distribution to more diverse interactions, including in-hand manipulation, tool use, clutter, and bimanual manipulation, and study how far the resulting priors can transfer to tasks that differ more substantially from pre-training. References [1] I. Akkaya, M. Andrychowicz, M. Chociej, M. Litwin, B. McGrew, A. Petron, A. Paino, M. Plappert, G. Powell, R. Ribas, et al. (2019) Solving rubik’s cube with a robot hand. arXiv preprint arXiv:1910.07113. Cited by: §1, §2, §3.2. [2] M. Andrychowicz, B. Baker, M. Chociej, R. Jozefowicz, B. McGrew, J. Pachocki, A. Petron, M. Plappert, G. Powell, A. Ray, J. Schneider, S. Sidor, J. Tobin, P. Welinder, L. Weng, and W. Zaremba (2020) Learning dexterous in-hand manipulation. International Journal of Robotics Research (IJRR) 39 (1), p. 3–20. Cited by: §2. [3] M. Bauza, J. E. Chen, V. Dalibard, N. Gileadi, R. Hafner, M. F. Martins, J. Moore, R. Pevceviciute, A. Laurens, D. Rao, M. Zambelli, M. Riedmiller, J. Scholz, K. Bousmalis, F. Nori, and N. Heess (2025) DemoStart: demonstration-led auto-curriculum applied to sim-to-real with multi-fingered robots. In IEEE International Conference on Robotics and Automation (ICRA), Cited by: §2. [4] T. Chen, M. Tippur, S. Wu, V. Kumar, E. Adelson, and P. Agrawal (2023) Visual dexterity: in-hand reorientation of novel and complex object shapes. Science Robotics 8 (84). Cited by: §1, §2, §4.1, Table 1. [5] T. Chen, J. Xu, and P. Agrawal (2022) A system for general in-hand object re-orientation. In Conference on Robot Learning (CoRL), Cited by: §2. [6] C. Cheng, M. Mukadam, J. Issac, S. Birchfield, D. Fox, B. Boots, and N. Ratliff (2018) RMPflow: a computational graph for automatic motion policy generation. In Workshop on the Algorithmic Foundations of Robotics (WAFR), Cited by: §2. [7] M. Dalal, M. Liu, W. Talbott, C. Chen, D. Pathak, J. Zhang, and R. Salakhutdinov (2025) Local policies enable zero-shot long-horizon manipulation. In IEEE International Conference on Robotics and Automation (ICRA), Cited by: §2. [8] A. T. Dang, J. Lee, M. Mukadam, X. A. Wu, B. Bucher, M. Nambi, and N. Fazeli (2026) HydroShear: hydroelastic shear simulation for tactile sim-to-real reinforcement learning. In Robotics: Science and Systems (RSS), Cited by: §2. [9] A. Handa, A. Allshire, V. Makoviychuk, A. Petrenko, R. Singh, J. Liu, D. Makoviichuk, K. Van Wyk, A. Zhurkevich, B. Sundaralingam, et al. (2023) Dextreme: transfer of agile in-hand manipulation from simulation to reality. In 2023 IEEE International Conference on Robotics and Automation (ICRA), p. 5977–5984. Cited by: §1, §2, §3.2. [10] J. Huang, Y. Ye, Y. Gong, X. Zhu, Y. Gao, and K. Zhang (2025) Spatially anchored tactile awareness for robust dexterous manipulation. arXiv preprint arXiv:2510.14647. Cited by: §2, §3.4. [11] M. Jaderberg, V. Dalibard, S. Osindero, W. M. Czarnecki, J. Donahue, A. Razavi, O. Vinyals, T. Green, I. Dunning, K. Simonyan, C. Fernando, and K. Kavukcuoglu (2017) Population based training of neural networks. arXiv preprint arXiv:1711.09846. Cited by: Appendix C, §3.2. [12] K. Kedia, T. G. W. Lum, J. Bohg, and C. K. Liu (2026) SimToolReal: an object-centric policy for zero-shot dexterous tool manipulation. In Robotics: Science and Systems (RSS), Cited by: §2, §2. [13] J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. Veness, G. Desjardins, A. A. Rusu, K. Milan, J. Quan, T. Ramalho, A. Grabska-Barwinska, et al. (2017) Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences 114 (13), p. 3521–3526. Cited by: §1, §2. [14] M. Laskin, D. Yarats, H. Liu, K. Lee, A. Zhan, K. Lu, C. Cang, L. Pinto, and P. Abbeel (2021) URLB: unsupervised reinforcement learning benchmark. In Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track, Cited by: §1, §2. [15] T. G. W. Lum, K. Kedia, C. K. Liu, and J. Bohg (2026) Play2Perfect: what matters in dexterous play pretraining for precise assembly?. arXiv preprint arXiv:2606.26428. Cited by: §2. [16] T. G. W. Lum, M. Matak, V. Makoviychuk, A. Handa, A. Allshire, T. Hermans, N. D. Ratliff, and K. Van Wyk (2024) DextrAH-G: pixels-to-action dexterous arm-hand grasping with geometric fabrics. In Conference on Robot Learning (CoRL), Cited by: §B.2, §B.3, §B.4, §B.4, §B.5, Appendix B, Appendix B, 2nd item, §1, §2, §2. [17] J. Luo, C. Xu, F. Liu, L. Tan, Z. Lin, J. Wu, P. Abbeel, and S. Levine (2024) FMB: a functional manipulation benchmark for generalizable robotic learning. International Journal of Robotics Research (IJRR). Cited by: 4th item, §1, §4.2, §4.2. [18] V. Makoviychuk, L. Wawrzyniak, Y. Guo, M. Lu, K. Storey, M. Macklin, D. Hoeller, N. Rudin, A. Allshire, A. Handa, and G. State (2021) Isaac gym: high performance GPU-based physics simulation for robot learning. In Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track, Cited by: §1. [19] M. Nakamoto, Y. Zhai, A. Singh, M. Sobol Mark, Y. Ma, C. Finn, A. Kumar, and S. Levine (2023) Cal-QL: calibrated offline rl pre-training for efficient online fine-tuning. In Neural Information Processing Systems (NeurIPS), Cited by: §1, §2. [20] A. Petrenko, A. Allshire, G. State, A. Handa, and V. Makoviychuk (2023) DexPBT: scaling up dexterous manipulation for hand-arm systems with population based training. External Links: 2305.12127, Link Cited by: §1, §3.2. [21] H. Qi, A. Kumar, R. Calandra, Y. Ma, and J. Malik (2022) In-hand object rotation via rapid motor adaptation. In Conference on Robot Learning (CoRL), Cited by: §2. [22] H. Qi, B. Yi, S. Suresh, M. Lambeta, Y. Ma, R. Calandra, and J. Malik (2023) General in-hand object rotation with vision and touch. In Conference on Robot Learning (CoRL), Cited by: §2. [23] N. D. Ratliff, J. Issac, D. Kappler, S. Birchfield, and D. Fox (2018) Riemannian motion policies. arXiv preprint arXiv:1801.02854. Cited by: §2. [24] N. D. Ratliff, K. Van Wyk, M. Xie, A. Li, and M. A. Rana (2020) Optimization fabrics. arXiv preprint arXiv:2008.02399. Cited by: §2. [25] S. Ross, G. Gordon, and D. Bagnell (2011) A reduction of imitation learning and structured prediction to no-regret online learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, p. 627–635. Cited by: §3.4. [26] J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov (2017) Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347. Cited by: §1, §3.1. [27] I. Shenfeld, J. Pari, and P. Agrawal (2025) RL’s razor: why online reinforcement learning forgets less. arXiv preprint arXiv:2509.04259. Cited by: §1, §2. [28] R. Singh, A. Allshire, A. Handa, N. Ratliff, and K. Van Wyk (2024) DextrAH-RGB: visuomotor policies to grasp anything with dexterous hands. arXiv preprint arXiv:2412.01791. Cited by: Appendix H, 2nd item, §2, §2, §3.4, §3.4. [29] L. Su, Z. Peng, R. Ren, S. Mao, J. Du, K. Zhang, and X. Zhu (2026) TacMap: bridging the tactile sim-to-real gap via geometry-consistent penetration depth map. arXiv preprint arXiv:2602.21625. Cited by: §2, §3.4. [30] K. V. Wyk, A. Handa, V. Makoviychuk, Y. Guo, A. Allshire, and N. D. Ratliff (2024) Geometric fabrics: a safe guiding medium for policy learning. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), Cited by: 2nd item, §2, §3.5. [31] P. Yin, T. Westenbroek, Z. Zhang, J. Tran, I. Dagnino, E. Shilamkar, N. Mbiziwo-Tiapo, S. Bagaria, X. Liu, G. Mullins, A. Kolobov, and A. Gupta (2026) Emergent dexterity via diverse resets and large-scale reinforcement learning. In The Fourteenth International Conference on Learning Representations, External Links: Link Cited by: §2. [32] Y. Zhao, R. Boney, A. Ilin, J. Kannala, and J. Pajarinen (2022) Adaptive behavior cloning regularization for stable offline-to-online reinforcement learning. In European Symposium on Artificial Neural Networks (ESANN), Cited by: §2. [33] Z. Zhou, A. Peng, Q. Li, S. Levine, and A. Kumar (2025) Efficient online reinforcement learning fine-tuning need not retain offline data. International Conference on Learning Representations (ICLR). Cited by: §2. Appendix A Task MDPs A.1 Reposing Task The total reward is: r=wr⋅exp(−αrdh)⏟hand-to-object alignment+wg⋅exp(−αgekp)⋅gc⏟object-to-goal alignment+wb⋅gc⏟contact gatingr= w_r· (- _r\,d_h)_hand-to-object alignment\;+\; w_g· (- _g\,e_kp)· g_c_object-to-goal alignment\;+\; w_b·\,g_c_contact gating (1) where: • Reward component weights: wr=1w_r=1, wg=5w_g=5, wb=0.01w_b=0.01. • αr=10 _r=10. dh=maxj‖jhand−‖2d_h= _j\|p_j^hand-p\|_2 is the maximum hand-to-object distance across all hand bodies. • αg _g: 15→3015→ 30 (ADR). ekpe_kp is the pose error, measured as the mean distance between 8 bounding-box keypoints of a cube of half-extent h=0.15h=0.15m of the current and goal object poses: ekp=18∑i=18‖i(,)−i(∗,∗)‖2e_kp= 18 _i=1^8\|K_i(p,q)-K_i(p^*,q^*)\|_2 (2) where (,)∈ℝ8×3K(p,q) ^8× 3 maps an object pose to its bounding-box corners in world frame. • gc∈0,1g_c∈\0,1\ is a contact gate active when the thumb and at least one other finger exert force above threshold τ=1τ=1N. • hq_h are hand joint positions and ¯h q_h is a nominal uncurled configuration. • Gravity is annealed from 00 to −9.81-9.81 m/s2 and object scale from 0.5–1.0×0.5--1.0× over 50 ADR increments, advancing when success rate exceeds 0.40.4. An episode succeeds when ekp<0.10e_kp<0.10 m. Episodes last 4 s. Domain Randomization (ADR) The following parameters are annealed from easy to hard over 50 ADR increments, triggered when the running success rate exceeds 0.4: • Gravity: 0→−9.810→-9.81 m/s2 (gravity annealing) • Fabric damping gain: 10→2010→ 20 • Goal sharpness: −15→−30-15→-30 (tighter precision required) Figure 5: Reposing Task. Visualizes the out-of-bound range in red, the goal sampling range in green, and the object spawn range in orange. (a) visualizes a frame of the initial grasp of the object and (b) visualizes a frame where the robot has completed the task by reaching the target pose frame. The object point clouds can be seen in green spheres, which are on the object. Primitive Objects for Pre-Training Figure 6: 16 primitive objects used during pre-training reposing task. The scale and physical properties of the objects are randomized during pre-training. A.2 Downstream Tasks from Scratch and Post-training All downstream manipulation tasks share a common reward function. Both training-from-scratch and bootstrapped policies use the same reward parameters; the only difference is initialization. The total reward is: r=wr⋅exp(−αrdh)⏟hand-to-object alignment+wg⋅exp(−αgekp)⏟object-to-goal alignmentr= w_r· (- _r\,d_h)_hand-to-object alignment\;+\; w_g· (- _g\,e_kp)_object-to-goal alignment (3) where: • Both regimes use identical reward parameters: wr=1w_r=1, αr=10 _r=10, wg=5w_g=5, αg=15 _g=15 (constant), with the finger-curl and lift terms disabled (wc=wℓ=0w_c=w_ =0). • Goal tolerance anneals from 0.05→0.020.05→ 0.02 m for both regimes over 50 ADR increments, advancing when success rate exceeds 0.40.4. An episode succeeds when epos<ϵgoale_pos< _goal. • dh=maxj‖jhand−‖2d_h= _j\|p_j^hand-p\|_2 is the maximum hand-to-object distance across all hand bodies. • hq_h are hand joint positions and ¯h q_h is a nominal uncurled configuration. • Δz=|pz−pz∗| z=|p_z-p_z^*| is the vertical distance between the object and its goal. • Reward component weights and sharpness values depend on the training regime: From scratch Bootstrap wrw_r, αr _r 11, 1010 11, 1010 wgw_g, αg _g 55, 1515 (constant) 55, 1515 (constant) • Goal tolerance anneals from 0.05→0.020.05→ 0.02 m over 50 ADR increments, advancing when success rate exceeds 0.40.4. An episode succeeds when epos<ϵgoale_pos< _goal. Figure 7: Downstream FMB Task. Visualizes the peg (orange) and board (sky blue) spawn ranges, object keypoints and goal pose keypoints. Grey keypoints represent the path set by the ADR level (denoted goal path annealing). At ADR 0, the goal is set to where the light blue sphere is placed and at ADR 50, the goal is set to the insertion goal where the light green sphere is placed. (a) shows sample initialization of the scene where the peg is dropped on to the table. In (b), we show how the reposing teacher reaches the goal at ADR 20, which is above the board. In (c), we visualize a sample policy rollout that inserts the peg into the board successfully where the keypoints align by the matching colors. The object point clouds in green spheres can be seen on the peg. A.3 ADEPT Pre-training and Post-training Algorithm Table 5 compares the PPO hyperparameters used for reposing pretraining and FMB post-training. Pretraining uses a higher actor learning rate, looser PPO clip, and adaptive learning-rate schedule to encourage exploration of the broad reposing task; post-training reduces both substantially and switches to a linear learning-rate decay to keep the adapted policy close to the pretrained behavior. The critic learning rate is held constant at 5×10−55× 10^-5 across both stages, and the discount factor γ and GAE τ are unchanged. Table 5: PPO hyperparameter comparison between reposing pretraining and FMB post-training. The actor learning rate is reduced by 100×100× and the PPO clip is tightened by 4×4× during post-training, while the critic learning rate is held constant. The actor learning-rate schedule changes from KL-adaptive (pretraining, with target KL = Actor KL threshold) to linear decay over Max epochs (post-training); the Actor KL threshold is logged but not used during post-training under the linear schedule. Hyperparameter Pretraining (reposing) Post-training (FMB) PPO Actor learning rate 1×10−31× 10^-3 1×10−51× 10^-5 Actor LR schedule KL-adaptive linear decay over Max epochs BC actor learning rate - 1×10−41× 10^-4 Max epochs 750,000750,000 200,000200,000 Critic learning rate 5×10−55× 10^-5 5×10−55× 10^-5 PPO clip ϵε 0.200.20 0.050.05 Actor KL threshold 0.0100.010 — Critic KL threshold 0.0160.016 0.0160.016 Discount factor γ 0.9980.998 0.9980.998 GAE τ 0.950.95 0.950.95 Mini-epochs (actor / critic) 5/55/5 4/44/4 Horizon length 3232 1616 Minibatch size 16,38416,384 24,57624,576 Entropy coefficient 1×10−31× 10^-3 2×10−42× 10^-4 Critic coefficient 44 44 Gradient-norm clip 1.01.0 1.01.0 Reward scale 0.010.01 0.010.01 Algorithm 1 Pretraining and Post-training for Bootstrapped Downstream Task Learning 1: Input: pretraining MDP ℳpreM_pre with obs opreo^pre, post-training MDP ℳpostM_post with obs opost≠opreo^post≠ o^pre (overlapping; neither is a superset), PPO clip ϵε 2: 3: Stage 1: reposing Teacher RL Training on ℳpreM_pre 4: Initialize pretrained policy πpre(⋅,ψ) _pre(·\,;ψ) and critic Vpre(⋅,ϕpre)V_pre(·\,; _pre) 5: for iteration =1,2,…=1,2,… do 6: Collect trajectories τ∼πpre(⋅∣otpre)τ _pre(· o_t^pre) in ℳpreM_pre 7: Update ψ,ϕpreψ, _pre with PPO 8: end for 9: 10: Stage 2: Actor Warm-Start via BC Distillation 11: Initialize downstream policy πpost(⋅,θ) _post(·\,;θ) 12: for iteration =1,2,…=1,2,… do 13: Collect trajectories τ∼πpost(⋅∣otpost)τ _post(· o_t^post) in ℳpostM_post 14: Query teacher actions: μpre,σpre←πpre(⋅∣otpre) _pre, _pre← _pre(· o_t^pre) 15: ℒBC←∑i(μposti(otpost)−μprei)2σprei 2+‖σpost(otpost)−σpre‖2L_BC← _i ( _post^i(o_t^post)- _pre^i)^2 _pre^i\,2+\| _post(o_t^post)- _pre\|_2 16: Update θ to minimize ℒBCL_BC 17: end for 18: 19: Stage 3: Frozen Actor Critic Warm-Up 20: Initialize critic Vpost(⋅,ϕ)V_post(·\,;φ), freeze θ 21: for iteration =1,2,…=1,2,… do 22: Collect trajectories τ∼πpost(⋅∣otpost)τ _post(· o_t^post) in ℳpostM_post, compute returns R^t R_t via GAE 23: Update ϕφ to minimize t[(Vpost(st)−R^t)2]E_t[(V_post(s_t)- R_t)^2] 24: end for 25: 26: Stage 4: RL Post-Training on ℳpostM_post 27: Unfreeze θ 28: for iteration =1,2,…=1,2,… do 29: Collect trajectories τ∼πpost(⋅∣otpost)τ _post(· o_t^post) in ℳpostM_post, compute R^t R_t, A^t A_t 30: Update θ,ϕθ,φ with PPO (clip ϵε) 31: end for 32: 33: Stage 5: Vision-Based Student Distillation 34: Initialize vision policy πω _ω 35: for iteration =1,2,…=1,2,… do 36: Collect trajectories τ∼πω(⋅∣otimg)τ _ω(· o_t^img) in ℳpostM_post with cameras 37: Query teacher actions: μpost,σpost←πpost(⋅∣otpost) _post, _post← _post(· o_t^post) 38: ℒDT←∑i(μωi(otimg)−μposti)2σposti 2+‖σω(otimg)−σpost‖2L_DT← _i ( _ω^i(o_t^img)- _post^i)^2 _post^i\,2+\| _ω(o_t^img)- _post\|_2 39: Update ω to minimize ℒDTL_DT 40: end for 41: 42: return πω _ω A.4 Observations per stage. Stage 1: reposing pretraining. PPO from scratch. Teacher (reposing policy) has privileged state; actor policy uses proprioceptive state + contacts + point cloud; critic adds privileged object state. No receptacle-specific observations. Stage 2: Bootstrap to downstream task. BC warm-start + conservative PPO. Initializes from the Stage 1 reposing actor. Adds receptacle pose and object–receptacle contact to both actor and critic. Critic keeps Stage 1’s privileged object state. Stage 3: Distillation (state → vision). BC from the Stage 2 state actor (now playing the teacher role). The new student replaces the point cloud with two RGB camera images (left and center). Table 6: Observation fields consumed at each pipeline stage in the current Cspace-mode codebase. Stage 1 = reposing pretraining (PPO, actor + asymmetric critic); Stage 2 = FMB post-training (PPO, actor + asymmetric critic); Stage 3 = two-camera RGB vision distillation, where a student network is supervised by the Stage 2 actor (treated as the teacher) via behavior cloning. Actor and student rows read the noisy variant of each proprioceptive field; the critic reads the corresponding clean variant. Each cell shows the dimensionality of the field when it is consumed at that stage; ✗ means the field is not consumed at that stage, and “image” denotes an image-tensor input. Rows newly introduced at Stage 2 (FMB extras / nominal delta) and Stage 3 (two-camera RGB) are highlighted. Stage 1: reposing Stage 2: FMB Stage 3: Distillation Observation field Actor Critic Actor Critic Teacher (actor) Student (actor) Proprioceptive / robot state robot_dof_pos (noisy) 23 23 23 23 23 23 robot_dof_vel (noisy) 23 23 23 23 23 23 hand_pos (5 bodies × 3) 15 15 15 15 15 15 hand_vel (5 bodies × 3) 15 15 15 15 15 15 actions 23 23 23 23 23 23 fabric_q, qd, qdd (cspace, 3×23) 69 69 69 69 69 69 fingertip_contacts (5 × 3) 15 15 15 15 15 ✗ nominal delta (joint + fingertip) ✗ ✗ ✗ 38 ✗ 38 Task / object state object_pos (noisy) 3 ✗ ✗ ✗ ✗ ✗ object_rot (noisy) 4 ✗ ✗ ✗ ✗ ✗ object state (pos + quat + vel) ✗ 13 ✗ 13 ✗ ✗ object_goal (position) 3 3 3 3 3 ✗ object_goal_quat 4 4 4 4 4 ✗ multi_object_idx_onehot 1 1 ✗ ✗ ✗ ✗ object_scale 1 1 ✗ ✗ ✗ ✗ Point cloud / vision pointcloud (64 pts × 3) 192 192 192 ✗ 192 ✗ RGB left camera (H×W×3H× W× 3) ✗ ✗ ✗ ✗ ✗ image RGB center camera (H×W×3H× W× 3) ✗ ✗ ✗ ✗ ✗ image Receptacle (FMB-specific) extras receptacle pose ✗ ✗ 7 7 7 ✗ object ↔ receptacle contact force ✗ ✗ 3 3 3 ✗ Privileged sim-only signals (critic only) hand body ang. vel. (5 bodies × 3) ✗ 15 ✗ ✗ ✗ ✗ palm force (first 3 of wrench) ✗ 3 ✗ ✗ ✗ ✗ palm wrench (full 6-D) ✗ ✗ ✗ 6 ✗ ✗ joint torque (measured) ✗ 23 ✗ 23 ✗ ✗ Flat dim total 391 438 392 280 392 206 ++ imgs Table 7: Flexiv–Sharpa (29 DoF). Observation fields consumed by the Rizon–Sharpa Cspace pipeline. Stage 1 uses the reposing PBT actor and asymmetric critic; Stage 2 uses the FMB bootstrap teacher; Stage 3 distills that state-based teacher into the stereo RGB + five-fingertip TacMap student. The entries and dimensions follow the checked-in agent configurations. “Image” denotes a tensor encoded outside the flat low-dimensional input. Stage 1: reposing Stage 2: FMB Stage 3: Distillation Observation field Actor Critic Actor Critic Teacher (actor) Student (actor) Proprioceptive / robot state proprio (29 joints + 6 FK bodies) 94 94 94 94 94 94 last_actions 29 29 29 29 29 ✗ fabric_q, qd, qdd (3×293× 29) 87 87 87 87 87 87 proprio_nominal_delta (29 joint + 18 FK) ✗ ✗ ✗ ✗ ✗ ✗ contact_forces_obs (5 fingers × 3) 15 15 15 15 15 ✗ contact_force_palm 3 3 3 3 3 ✗ Task / object state object_pos_noisy 3 ✗ ✗ ✗ ✗ ✗ object_rot_noisy 4 ✗ ✗ ✗ ✗ ✗ object_state (pos + quat + velocity) ✗ 13 ✗ 13 ✗ ✗ object_goal 3 3 3 3 3 ✗ object_goal_quat 4 4 4 4 4 ✗ multi_object_idx_onehot 1 1 ✗ 1 ✗ ✗ object_scale 1 1 ✗ 1 ✗ ✗ Point cloud / RGB / tactile perception pc_obs (64 points × 3) 192 192 192 192 192 ✗ stereo RGB (left + right) ✗ ✗ ✗ ✗ ✗ 2 images sharpa_tactile_deform ✗ ✗ ✗ ✗ ✗ 5 images sharpa_tactile_anchor_xyz (5×35× 3) ✗ ✗ ✗ ✗ ✗ 15 FMB receptacle extras board_pose ✗ ✗ 7 7 7 ✗ object_target_contact_force ✗ ✗ 3 3 3 ✗ Privileged sim-only critic signals hand_forces (palm wrench) ✗ 6 ✗ 6 ✗ ✗ measured_joint_torque ✗ 29 ✗ 29 ✗ ✗ Flat dim total 436 477 437 487 437 196 + images Appendix B Joint Configuration Space Geometric Fabric Setup This appendix details the geometric fabric components, parameters, and policy–fabric integration loop summarized in Sec. 3.5. We adopt the notation of [16]: the fabric state f,˙f,¨f∈ℝnqq_f, q_f, q_f ^n_q evolves according to f(f,˙f)¨f+f(f,˙f)+π(t)= 0,M_f(q_f, q_f)\, q_f\;+\;f_f(q_f, q_f)\;+\;f_π(a_t)\;=\;0, (4) where f≻0M_f 0 is the system metric (mass) assembled from all fabric components by pull-back, ff_f is the nominal path-generating geometric force, and πf_π is the action-driven forcing term. The policy emits a cspace target t∗=(t)∈ℝnqq^*_t= π(a_t) ^n_q at 60Hz60\,Hz; the fabric is forward-integrated with the second-order Runge–Kutta scheme of [16] at 60Hz60\,Hz. For the 23 DoF KUKA-Allegro system (nq=7arm+16fingern_q=7\;arm+16\;finger) and the 29 DoF Flexiv-Sharpa system (nq=7arm+22fingern_q=7\;arm+22\;finger), we partition the cspace target into arm and hand slices, t∗=[t∗,a;t∗,h]q^*_t=[q^*,a_t;\,q^*,h_t] with t∗,a∈ℝ7q^*,a_t ^7 and t∗,h∈ℝ16q^*,h_t ^16 or t∗,h∈ℝ22q^*,h_t ^22 respectively. B.1 Pull-Back from Taskmaps to the Root Every fabric component lives on a taskmap =ϕ(f)x= φ(q_f) with Jacobian =∂ϕ/∂fJ=∂ φ/ _f and contributes a leaf metric (,˙)M(x, x) and leaf force (,˙)f(x, x). Root-space contributions are f+=⊤,f+=⊤(+˙f),M_f\; +=\;J M\,J, _f\; +=\;J \! (f\;+\;M\, J\, q_f ), (5) the second term being the curvature force required to keep the leaf acceleration equal to the pulled-back leaf force under the chain rule. B.2 Cspace Attractors For each body part p∈a,hp∈\a,h\ (arm and hand), the cspace taskmap is the slice p=pfx^p=S_pq_f with pS_p the constant index-selector. The position error is p=p−t∗,pe^p=x^p-q^*,p_t. Mass. The metric is isotropic with a smooth switch between a minimum and maximum mass around the error norm: mp(p)=(mmaxp−mminp)σ(−σmp(‖p‖−δmp))+mminp,p(p)=mp(p),m^p(e^p)\;=\;(m^p_ -m^p_ )\,σ\! (- _m^p (\|e^p\|- _m^p ) )\;+\;m^p_ , ^p(e^p)\;=\;m^p(e^p)\,I, (6) where σ(z)=12(tanhz+1)σ(z)= 12( z+1), σmp _m^p is the mass sharpness and δmp _m^p the mass switch offset. By default, mminp=mmaxpm^p_ =m^p_ , so the switch is inert and p=mpM^p=m^pI. Forcing attractor. A homogeneous-degree-one (HD1) acceleration pulls px^p toward t∗,pq^*,p_t with conical shaping along the error direction and a soft attraction-radius damping: ¨forcp x^p_forc =−kaptanh(αap‖p‖)p‖p‖−bp(p)˙p, =-\,k^p_a\, \! (α^p_a\|e^p\| )\, e^p\|e^p\|\;-\;b^p(e^p)\, x^p, (7) bp(p) b^p(e^p) =bmaxpσ(−σdp(‖p‖−rdp)), =b^p_ \,σ\! (- _d^p (\|e^p\|-r_d^p ) ), (8) with conical gain kapk^p_a, conical sharpness αapα^p_a, damping ceiling bmaxpb^p_ , damping sharpness σdp _d^p, and damping radius rdpr_d^p. The damping engages only inside the attraction radius rdpr_d^p, providing critical damping near the target without slowing long approaches. Geometric attractor. In parallel with the forcing term we attach a homogeneous-degree-two (HD2) geometric attractor with the same direction but quadratic-in-velocity scaling ¨geomp=−kap,g‖˙p‖2tanh(αap,g‖gp‖)gp‖gp‖, x^p_geom\;=\;-\,k^p,g_a\,\| x^p\|^2\, \! (α^p,g_a\|e^p_g\| )\, e^p_g\|e^p_g\|, (9) where gp=p−postpe^p_g=x^p-q^p_post and postpq^p_post is a config-driven default that postures the elbow out and fingers curled. Since ¨geomp x^p_geom is HD2 in ˙p x^p, it produces speed-invariant paths and does not bias convergence to t∗,pq^*,p_t in steady state [16]. B.3 Body Sphere Collision Avoidance We approximate the robot geometry with Ns=31N_s=31 collision spheres covering the arm links, palm, knuckles, and finger phalanges. Forward kinematics maps fq_f to sphere origins s(i)=ϕfk,i(f)x^(i)_s= φ_fk,i(q_f) with Jacobian s(i)∈ℝ3×nqJ^(i)_s ^3× n_q. For each sphere–obstacle pair (environment meshes and declared self-collision sphere–sphere pairs), let rir_i be the sphere radius, did_i the signed clearance, ^i=(i−s(i))/‖i−s(i)‖ n_i=(r_i-x^(i)_s)/\|r_i-x^(i)_s\| the unit direction to the nearest obstacle point ir_i, and ˙s(i) x^(i)_s the sphere velocity. The signed closing speed is vi=−^i⋅(˙s(i)−iobs),v_i\;=\;-\, n_i·( x^(i)_s-v^obs_i), (10) with iobsv^obs_i the velocity of the closest point on the obstacle (rigid composition of linear and angular velocity, supporting moving obstacles). Following [16], define si s_i =12(tanh(−α1(vi−α2))+1),(smooth velocity gate) \;=\; 12\! ( \! (- _1(v_i- _2) )+1 ), (smooth velocity gate) (11) d¯i d_i =max(dmin,di), \;=\; (d_ ,\,d_i), (12) where sis_i vanishes when the sphere retreats and saturates to 11 when it approaches. The leaf metric and leaf acceleration for sphere i are b(i)=sid¯i 3^i⊗^i,¨b(i)=sid¯i 3^i.M^(i)_b\;=\; s_i d_i^\,3\; n_i n_i, x^(i)_b\;=\; s_i d_i^\,3\; n_i. (13) A per-sphere Frobenius normalization preserves Eigen-directions while making the response invariant to absolute distance, ^b(i)=b(i)/‖b(i)‖F M^(i)_b=M^(i)_b/\|M^(i)_b\|_F, and the resulting block is rescaled by a budget-allocated weight that caps the aggregate stiffness across all spheres, wi=κ(ri/‖2)d¯i 2,w~i=min(∑jwj,ℬ)wi∑jwj,~b(i)=w~i^b(i),w_i\;=\; κ\,(r_i/\|r\|_2) d_i^\,2, w_i\;=\; \! ( _jw_j,\;B )\,w_i _jw_j, M^(i)_b\;=\; w_i\, M^(i)_b, (14) where κ is a metric scalar and ℬB is the total metric budget (separate values for forcing vs. geometric channels). The leaf acceleration is split into a forcing and a geometric component: ¨b,forc(i) x^(i)_b,forc =−kf^i−bc(˙s(i)−iobs), \;=\;-\,k_f\, n_i\;-\;b_c\,( x^(i)_s-v^obs_i), (15) ¨b,geom(i) x^(i)_b,geom =−kg‖˙s(i)−iobs‖2^i. \;=\;-\,k_g\,\| x^(i)_s-v^obs_i\|^2\, n_i. (16) Each is pulled back through its own sphere Jacobian and summed: collf=∑i(i)⊤s~(i)b(i)s,collf=−∑i(i)⊤s~(i)b¨(i)b.M^coll_f\;=\; _iJ^(i) _s\, M^(i)_b\,J^(i)_s, ^coll_f\;=\;- _iJ^(i) _s\, M^(i)_b\, x^(i)_b. (17) Per-sphere pull-back (rather than stacked-vector pull-back) ensures that the budget weighting acts on the unit-Frobenius blocks ^b(i) M^(i)_b rather than on the cross-sphere normalization, eliminating dilution as NsN_s grows. B.4 Joint Limit Repulsion Two taskmaps, one per limit side, are defined for each joint i: xiU=q¯i−qf,iq¯i−q¯i,xiL=qf,i−q¯iq¯i−q¯i,xiU,xiL∈[0,1],x^\,U_i\;=\; q_i-q_f,i q_i- q_i, x^\,L_i\;=\; q_f,i- q_i q_i- q_i, x^\,U_i,x^\,L_i∈[0,1], (18) i.e. the signed clearance is normalized by the joint range, so all tuning parameters live in “fraction of range” units regardless of physical units (rad vs. m for revolute vs. prismatic joints). The Jacobian rows are ±diag(1/(q¯i−q¯i)) (1/( q_i- q_i) ). The leaf metric and acceleration on each side are J(,˙) ^\,J(x, x) =diag(gi(x˙i)kb(max(xi−δJ,ϵ))2), \;=\;diag\! (\,g_i( x_i)\, k_b( (x_i- _J,\,ε))^2\, ), (19) gi(x˙i) g_i( x_i) =σ(−σg(x˙i−δg))∈[0,1], \;=\;σ\! (- _g( x_i- _g) )∈[0,1], (20) ¨forcJ x^\,J_forc =kr−bJ 1˙<0⊙˙, \;=\;g\,k_r\;-\;b_J\,1\ x<0\ x, (21) ¨geomJ x^\,J_geom =‖˙‖2kr, \;=\;\| x\|^2\,k_r\,g, (22) where kbk_b is the metric scalar, δJ _J the metric-exploder offset (the barrier saturates within the last δJ _J of the range), ϵε clamps the metric below a configured maximum, krk_r is the soft-ReLU repulsion gain, and bJb_J a damping gain that engages only when x˙i<0 x_i<0. The gate gig_i is a smooth tanh switch (replacing the bang-bang gate of [16]) that tapers the barrier when the joint retreats from the limit, suppressing chatter at the approach/retreat boundary while preserving the unbounded 1/(x−δJ)21/(x- _J)^2 stiffness as x→δJx→ _J. Per-joint acceleration and jerk caps. On top of the JL barrier, a post-hoc quadratic-program step [16] rescales the integrated acceleration to respect per-joint caps, q¨f,imax=min(q¨imax,Δt⋅q˙imax2q¨imax), q_f,i^\, \;=\; \! ( q_i^\, ,\; t· q_i^\, 2\, q_i^\, ), (23) with per-joint values q¨amax=7.5, 7.5, 10, 10, 10, 20, 20rad/s2 q^\, _a=\7.5,\,7.5,\,10,\,10,\,10,\,20,\,20\\,rad/s^2 for the arm and q¨hmax=22.5rad/s2 q^\, _h=22.5\,rad/s^2 for all sixteen hand joints (jerk caps scale accordingly). B.5 Cspace Damping and Speed Control Two velocity-dependent terms regulate end-to-end motion smoothness and speed. The first is a constant cspace damping applied to the mass-momentum f˙fM_f q_f: fdamp=kdf˙f,kd=10(exposed to ADR).f^\,damp_f\;=\;k_d\,M_f\, q_f, k_d=10\;\;(exposed to ADR). (24) The second is a speed-control term that caps the full system kinetic-energy level E⋆E . We attach Euclidean energies to two taskmaps: (i) a slice h¯=h¯fx h=S_ hq_f over the non-hand joints (here, the 7 arm joints) with scale βh¯=0.75 _ h=0.75, and (i) a robot-frame-origins taskmap palm∈ℝ3x^palm ^3 over the palm origin with scale βpalm=0.25 _palm=0.25. Each contributes E(k)=βk,E(k)=12βk˙(k)⊤˙(k),k∈h¯,palm.M_E^(k)= _k\,I, E^(k)\;=\; 12\, _k\, x^(k) x^(k), k∈\ h,palm\. (25) Hand joints are deliberately excluded so finger motion is decoupled from the energization budget (finger joints are free to move as fast as fabric terms allow). The pulled-back energy metric and total energy are E=∑k(k)⊤E(k)(k),E(f,˙f)=∑kE(k).M_E\;=\; _kJ^(k) \,M_E^(k)\,J^(k), E(q_f, q_f)\;=\; _kE^(k). (26) The energization coefficient [16] that holds E constant under the current geometric force is α=−˙f⊤(E¨f,geom+E)˙f⊤E˙f+ε,¨f,geom=−f−1f,α\;=\;-\, q_f^\, \! (M_E\, q_f,geom+f_E ) q_f^\, \,M_E\, q_f+ , q_f,geom\;=\;-M_f^-1\,f_f, (27) with E=f_E=0 for Euclidean energies. The energized force is f⋆=f−αf˙ff_f =f_f-α\,M_f\, q_f, which annihilates the velocity component along the geometric flow. Finally, the speed-control damping engages only when the current energy exceeds the target: fsc=bsc 1E>E⋆f˙f,E⋆=1.0,bsc=100(both exposed to ADR).f^\,sc_f\;=\;b_sc\,1\E>E \\;M_f\, q_f, E =1.0,\;\;b_sc=100\;\;(both exposed to ADR). (28) Equations (24)–(28) yield a trajectory whose end-to-end speed is consistent across configurations: long motions are energized to the target, short motions are passively damped. B.6 Action-to-Cspace Mapping The policy outputs t∈[−1,1]nqa_t∈[-1,1]^n_q with nq=23n_q=23 for the KUKA-iiwa7 + Allegro system (77 arm + 1616 finger joints), interpreted as per-joint relative deltas scaled by δmax=0.1 _ =0.1 rad. The Cspace target that enters the policy forcing term πf_π is t∗=f(t−1)+δmaxtq^*_t=q_f^(t-1)+ _ \,a_t, clamped to the URDF joint limits, where f(t−1)q_f^(t-1) is the fabric’s integrated joint configuration from the previous step. This relative formulation keeps each policy step bounded in magnitude and avoids the large position jumps that an absolute target would allow, while still letting the policy exploit the full 2323-DoF action manifold over the course of a trajectory. B.7 Policy–Fabric Integration Loop At each control step, the action-mapped Cspace target t∗q^*_t is broadcast to the fabric. The fabric performs K inner integration steps of the fabric ODE with timestep Δt t via a displacement integrator (default K=2K=2), producing (f,˙f,¨f)(q_f, q_f, q_f). The integrated position and velocity are then issued to the underlying joint controller (PD in simulation, an admittance controller on the real robot). The fabric is captured into a CUDA graph so the per-step cost remains negligible at the parallel scale required for large-batch RL. B.8 ADR-Annealed Fabric Parameters Several fabric parameters are part of the ADR curriculum rather than fixed at YAML defaults: • Cspace damping gain. Annealed from a high value (heavily smoothed, easy to control) toward a lower value (more reactive) as success rate improves. • Speed-control energy target. Annealed to broaden the range of motion speeds the policy is exposed to. • Velocity attenuation. A scalar applied between the fabric velocity output and the joint controller, also annealed via ADR. The fabric is therefore not treated as a fixed black-box controller but as part of the curriculum: as policy competence grows, the regularization the fabric provides is gradually relaxed. Appendix C Population-Based Training We train the reposing pretraining teacher with decentralized Population-Based Training (PBT) [11] to efficiently search the joint PPO + ADR hyperparameter space. Ranking and replacement. We launch N=16N=16 policies in parallel. After a warm-up of Tstart=200MT_start=200M environment frames, every policy independently performs a PBT check every TPBT=200MT_PBT=200M frames: it reads the latest metadata files from all other policies, restricts the comparison to checkpoints from policies that have collected approximately the same number of frames (so newer policies are not penalized for having less data), and ranks the population by target objective. The bottom rreplace=0.4r_replace=0.4 fraction of the population is eligible for replacement: a bottom-quartile policy within a small fractional-standard-deviation and absolute threshold of the population leader only has its hyperparameters mutated, while one significantly behind the leader additionally has its policy weights overwritten with those of a uniformly-randomly-selected top-rreplacer_replace policy. ADR is preserved on replacement. When a low-performing policy’s weights are rewritten from a top performer, it continues training from its own ADR level rather than the donor’s. Because the receiving policy typically has a lower ADR level than the donor, the rewritten policy effectively resumes training against an easier curriculum slice but with the donor’s well-shaped weights, letting it climb back toward the population frontier without re-discovering the basic dexterous behaviors the donor already encodes. Mutation. At every PBT check, each of the following PPO hyperparameters is independently re-sampled with probability pmut=0.25p_mut=0.25 via a log-uniform perturbation (multiplied or divided by a factor sampled uniformly in [1.1,2.0][1.1,2.0]): • Learning rate, gradient-norm clip, entropy coefficient, critic coefficient, bounds-loss coefficient, and KL threshold (all via mutate_float). • PPO clip ϵε via mutate_eps_clip, clamped to [0.01,0.3][0.01,0.3]. • Number of PPO mini-epochs via mutate_mini_epochs, clamped to [1,12][1,12] with ±1± 1–33 integer steps. • Discount factor γ and GAE τ via mutate_discount, which mutates 1−x1-x rather than x to be conservative on values close to 11. Target objective. Policies are ranked by their running success rate at their current ADR level, exposed to PBT via the simulator’s true_objective info field rather than the shaped reward. Using the unshaped success rate as the ranking signal means PBT compares policies on actual task progress rather than on shaped-reward magnitudes that fluctuate with mutated PPO hyperparameters (e.g. entropy coefficient, value loss scale) and with each policy’s ADR level. Appendix D Post-Training Recipes This appendix documents the practical recipe choices that make the structured RL post-training of Sec. 3.3 work in practice. PPO hyperparameter values are in Tab. 5; the items below are the higher-level knobs that govern Stages 2–4 of Alg. 1. Stage 2 — Actor warm-start via BC distillation. • BC for 40,00040,000 iterations is sufficient to converge the student’s action distribution to the teacher’s on the downstream observation space. • BC rollouts use a mixed student-teacher policy that alternates between teacher and student actions, which keeps the BC dataset on the states the student will visit at deployment while remaining anchored on teacher-reachable states. • The critic is not updated during BC; only the actor minimizes the Mahalanobis-weighted distance to the teacher action distribution. • BC is pinned at ADR step 20 (Sec. 4.1), where the pretrained teacher zero-shots reliably: the goal is the single ADR 20 pose on the lift-and-transport segment, and ADR advancement is disabled during BC. The goal-path annealing flag is kept on so the student tracks the teacher through the lift-and-transport segment. Stage 3 — Critic warm-up with frozen actor. • From this stage onward, goal-path annealing is turned off and the goal is set to the final insertion endpoint at ADR 50; the ADR randomizations continue to advance. For the first 2020 post-BC epochs the actor is frozen and only the critic updates against rpostr_post. This aligns value estimates with ℳpostM_post before any policy-gradient step modifies the actor. • During this window the actor’s log-std is fixed at σ=−2σ=-2 (≈0.14≈ 0.14 in linear scale), keeping exploration narrow around the BC-distilled mean so the critic targets stay on-distribution. • The critic learning rate is 5×10−55× 10^-5 throughout (unchanged from Stage 4). Stage 4 — Conservative PPO post-training. • Actor learning rate drops to 1×10−51× 10^-5 (100×100× lower than pretraining) with linear decay over 200,000200,000 epochs; PPO clip tightens to ϵ=0.05ε=0.05 (4×4× tighter) to bound each policy update. • Goal-path annealing is turned off in PPO: the goal sits at the actual insertion endpoint so PPO learns the contact-rich insertion itself, while ADR can continue to advance. • The post-training actor–critic uses separate actor and critic trunks (rather than the shared-trunk architecture of pretraining), allowing the critic to specialize on rpostr_post without contaminating actor features. • Observation clipping is relaxed from 5.05.0 (default) to 100.0100.0 to accommodate the larger magnitudes of receptacle pose and object–receptacle contact-force observations introduced at post-training. Task-specific reward shaping for FMB peg insertion. • Both peg–board contact penalty and receptacle contact penalty are set to 00. The pretrained policy already handles contacts gracefully, and insertion explicitly requires sustained contact between the peg, fingertips, and receptacle. • External object wrenches are disabled during post-training, since the peg is grasped throughout. • Peg orientation randomization is set to its maximum range ([1.0,1.0][1.0,1.0]) from the start of post-training rather than ramped via ADR, and the object-to-goal reward sharpness is fixed at 1515 rather than annealed. Appendix E ADEPT Training Efficiency Figure 8: Training curves on FMB peg insertion comparing (1) training from scratch, (2) direct PPO finetuning of the pretrained reposing teacher, and (3) ADEPT post-training. Blue inset: success-rate collapse for (2) during transfer. Figure 9: Training curves across different seeds for training from scratch. ADEPT post-training reliably converges to a solution while most seeds for training from scratch either plateau or take a very long time to train for a fixed compute budget. Appendix F Real-World Deployment Topology Figure 10: Real-world camera views for (a) Kuka-Allegro and (b) Flexiv-Sharpa setups. The real-world deployment of ADEPT preserves the exact action interface used in simulation. The policy commands a Cspace target through the geometric fabric and receives observations from a graph identical to the one used during training; the only substantive change is that the C++ admittance controller and RealSense driver replace the sim physics and renderer. Camera transport. The two Intel RealSense cameras (one mounted to the left of the workspace, one centered above the workbench) are read by a C++ camera driver that resizes the images on-board to 320×240320× 240 and publishes them to the policy process over ZMQ. Compute split and processes. The control stack is split across two machines. An NVIDIA Jetson AGX Orin sits next to the robot and hosts (i) a real-time C++ admittance controller that integrates the fabric-derived mass-and-force pair (f,f+π)(M_f,f_f+f_π) from the fabric ODE in Sec. 3.5 into joint position and velocity commands, and (i) the Python fabric process that evaluates the geometric fabric of Sec. 3.5. The student policy runs on a separate workstation with a discrete GPU and receives proprioception and the two RGB streams over ZMQ; it issues 23-D Cspace targets back to the fabric, which integrates them into the smooth joint trajectories consumed by the admittance controller. Joint state and the (f,f+π)(M_f,f_f+f_π) pair are exchanged between the C++ controller and the fabric process through POSIX shared memory on the Jetson. Rate hierarchy. The C++ admittance controller runs at 1 kHz, with the underlying KUKA control loop at 1 kHz and the Allegro control loop at 333 Hz. All Python-side processes (fabric, sensor, policy) share a common 6060 Hz cadence, while the high-rate hardware loops run underneath. Appendix G Distillation Loss Details This appendix gives the full forms of the behavior-cloning and auxiliary losses that constitute the distillation objective ℒ=ℒBC+ℒauxL=L_BC+L_aux introduced in Sec. 3.4. The student policy πθ(at∣ot) _θ(a_t o_t) is a stochastic Gaussian with mean μθ(ot) _θ(o_t) and standard deviation σθ(ot) _θ(o_t), and the teacher πT(at∣otT) _T(a_t o_t^T) has mean μT(otT) _T(o_t^T) and fixed standard deviation σT _T. G.1 Behavior-Cloning Loss The BC term matches both moments of the student’s action distribution to the teacher’s via a Mahalanobis-weighted L2 distance, ℒBC(ot,otT,θ)=∑i=1nq(μθi(ot)−μTi(otT))2(σTi)2+∑i=1nq(σθi(ot)−σTi)2(σTi)2,L_BC(o_t,o_t^T;\,θ)= _i=1^n_q ( _θ^i(o_t)- _T^i(o_t^T) )^2( _T^i)^2\;+\; _i=1^n_q ( _θ^i(o_t)- _T^i )^2( _T^i)^2, (29) where nq=23n_q=23 for the KUKA-iiwa7 + Allegro system. The teacher’s fixed std-dev σT _T serves as a per-dimension precision weighting: joints the teacher is more confident about (smaller σT _T) contribute more strongly to the loss. G.2 Auxiliary 8-Keypoint Pose Loss The auxiliary head predicts the eight bounding-box corners of the peg in the world frame, and the loss is the L2 norm of the per-corner residuals, ℒaux(ot,θ)=∑i=18‖^θi(ot)−t⋆,i‖22,L_aux(o_t;\,θ)= _i=1^8 \| k^i_θ(o_t)-k ,i_t \|_2^2, (30) where ^θi(ot)∈ℝ3 k^i_θ(o_t) ^3 is the student’s predicted 3D world-frame position of the i-th corner, and t⋆,i∈ℝ3k ,i_t ^3 is the corresponding ground-truth corner position obtained by rotating the eight half-extent offsets of the peg by the simulator’s peg orientation and adding the peg position. G.3 Total Distillation Objective Combining the two terms with their respective weighting coefficients, ℒdistill(θ)=ot∼πθ[wBC⋅ℒBC(ot,otT,θ)+waux⋅ℒaux(ot,θ)],L_distill(θ)=E_o_t _θ\! [\,w_BC·L_BC(o_t,o_t^T;\,θ)\;+\;w_aux·L_aux(o_t;\,θ)\, ], (31) where wBCw_BC and wauxw_aux are scalar weighting coefficients and the expectation is over rollouts ot∼πθo_t _θ collected by the student in the same task environment. We use wBC=1w_BC=1 and waux=20w_aux=20. Appendix H Visuo-Tactile Student Architecture This appendix details the visuo-tactile student used on the Flexiv-Sharpa embodiment. The vision-only KUKA-Allegro student is a strict subset of the same architecture with the tactile branch removed. Observations. The student consumes the two-camera RGB pair (tL,tC)(I^L_t,I^C_t) from left and center cameras, noisy proprioception tq_t, fabric state tfq^f_t, per-finger TacMap depth maps tkk=15\T^k_t\_k=1^5, and per-finger fingertip positions tkk=15∈ℝ3\x^k_t\_k=1^5 ^3 obtained via forward kinematics on the noisy joint state. Vision encoder. The two-camera pair is passed through a shared ResNet backbone followed by a cross-attention fuser [28] to a dvisd_vis-dimensional vision latent (dvis=256d_vis=256). The backbone is initialized from the Stage-1 pretrained student (Sec. 3.4) and unfrozen on the DextrAH-RGB [28] schedule. Tactile encoder. Each per-finger depth map tk∈ℝH×WT^k_t ^H× W is stacked with its thresholded binary contact map [tk≥τ]1[T^k_t≥τ] (τ=1/255τ=1/255) into a two-channel input, and passed through a small CNN (two 3×33×3 convolutions with stride 2 and channel widths 16 then 32, adaptive average pooling, and a linear projection with LayerNorm and ELU) shared across fingers. The CNN emits a per-finger feature tk∈ℝdfu^k_t ^d_f with df=32d_f=32. FiLM spatial anchoring. Each fingertip position tkx^k_t is expanded with B=4B=4 Fourier bands, γ()=[,sin(20π),cos(20π),…,sin(2B−1π),cos(2B−1π)]γ(x)=[x,\, (2^0 ), (2^0 ),…, (2^B-1 ), (2^B-1 )], and mapped through a two-layer MLP (hidden 128, ELU) to per-finger FiLM parameters (tk,tk)∈ℝdf×ℝdf( γ^k_t, β^k_t) ^d_f×R^d_f. The modulated per-finger feature is ~tk=(1+αtk)⊙tk+αtk, u^k_t=(1+α γ^k_t) ^k_t+α β^k_t, (32) with α=0.1α=0.1 scaling the modulation to keep it a soft perturbation on the shared per-finger CNN features rather than a hard override. The five modulated per-finger features are flattened into a tactile latent t∈ℝdtacu_t ^d_tac with dtac=5⋅df=160d_tac=5· d_f=160. Fusion and policy trunk. The vision, tactile, and proprioception latents are concatenated with the fabric state and passed through a two-layer MLP with widths [512,512][512,512] and ELU activations, followed by a 1024-unit LSTM (layer-normalized, placed before the final linear heads). The final head emits a mean action θ(ot)∈ℝnq μ_θ(o_t) ^n_q with a fixed log-standard-deviation, matching the teacher’s action parameterization. Training objective. The student is trained with pure teacher BC using the objective in Appx. G, with a single auxiliary term supervising an 8-keypoint peg pose prediction from the shared vision features. On Flexiv-Sharpa FMB the aux term uses a peg-tight-keypoints target with coefficient waux=20w_aux=20; other auxiliary heads (object position, object 6-D rotation, board pose, per-finger contact force and binary contact) are exposed by the codebase but disabled for this experiment. The BC target is filtered through a soft z-mask that down-weights teacher actions in low-height regions of the state distribution (threshold 0.080.08 m, softness 0.020.02, floor 0.10.1) to avoid over-fitting to teacher table-clearance behavior that the real robot cannot exactly reproduce. Appendix I Distillation Domain Randomization To enable zero-shot sim-to-real transfer, both stages of the teacher–student distillation curriculum (Sec. 3.4) apply aggressive domain randomization across physics, scene layout, sensor noise, and visual rendering. The randomization schedule for each stage matches the corresponding RL training stage so that the student observes the same environmental distribution that the teacher was trained on. Table 8 summarizes the full set of parameters and their ranges; the values reported correspond to the Stage 2 (downstream FMB) regime, with the analogous Stage 1 ranges differing only in scene layout (single fixed receptacle, full SO(3)SO(3) peg orientation). Figure 11: A sample 16x16 grid of visual randomizations on the left camera of the FMB training scene for distillation. Table 8: Domain randomization parameters used during teacher–student distillation. Per-env ranges are sampled uniformly at episode reset; ADR-annealed ranges advance from the lower bound to the upper bound over training as the running success rate exceeds the ADR threshold; per-step noise is added at every control step; per-frame randomization is applied at each camera render. Parameter Range / Distribution Physics (per-env at reset, ADR-annealed) Object mass scale (0.5, 3.0)×(0.5,\,3.0)× nominal Object static friction (0.5, 1.2)(0.5,\,1.2) Object dynamic friction (0.3, 1.0)(0.3,\,1.0) Object restitution (0.8, 1.0)(0.8,\,1.0) Robot joint stiffness scale (0.5, 2.0)×(0.5,\,2.0)× default Robot joint damping scale (0.5, 2.0)×(0.5,\,2.0)× default Robot joint friction (0, 5.0)U(0,\,5.0) Nm Scene and object spawn (per-env at reset, ADR-annealed) Peg spawn position (x,y)(x,y) jitter up to ±15± 15 cm per axis Peg spawn orientation full SO(3)SO(3) at final ADR Random wrench on peg up to 1010 m/s2 linear acceleration equivalent Receptacle / board pose jitter max range [−7,12][-7,12] cm in x, [−30,10][-30,10] cm in y Proprioception and object-state noise (per-step on observations) Joint position Gaussian noise (0, 0.082)N(0,\,0.08^2) rad Joint position bias (per-env) (−0.08, 0.08)U(-0.08,\,0.08) rad Joint velocity Gaussian noise (0, 0.182)N(0,\,0.18^2) rad/s Joint velocity bias (per-env) (−0.08, 0.08)U(-0.08,\,0.08) rad/s Object position Gaussian noise (0, 0.032)N(0,\,0.03^2) m Object orientation Gaussian noise (0, 0.12)N(0,\,0.1^2) rad Visual randomization (per-frame at camera rendering) Lighting (intensity, color, direction) random per env Dome / background texture random from texture library Table material random color and roughness Robot link materials random color and roughness Peg diffuse tint (RGB) per-channel (0.90, 1.0)U(0.90,\,1.0) (near-white) Board diffuse tint (RGB) per-channel (0.04, 0.12)U(0.04,\,0.12) (near-black) Camera position jitter ±3± 3 cm from calibrated pose Camera rotation jitter ±3∘± 3 from calibrated pose Appendix J Qualitative Analysis of Emergent Grasping Behaviors Fig. 12 highlights representative grasp modes from each training condition. Panels (e, f) show the reposing teacher applied zero-shot at ADR level 20: the grasps look natural, with the fingers wrapping around the peg in multi-contact closure, but they are not always task-aligned. A common failure mode is grasping the peg bottom-up with the hand below the peg as shown in (f). This works for the reposing segment but cannot complete insertion, which requires grasping the peg from above with its legs pointing down toward the hole. Panels (g) and (h) show that ADEPT post-training removes the bottom-up grasps entirely while preserving the natural finger-wrap behavior inherited from reposing pretraining. Panels (a-d) show that policies trained from scratch do not develop the natural finger-wrap behavior at all; the grasps are functional for insertion but lack the multi-contact stability that reposing pretraining produces. For example, (a) and (c) grasp the peg using only the index and pinky fingers, a configuration that is highly unusual for precise human manipulation. In (b) and (d), the fingers are curled to their extreme with just one finger wrapping around the peg to stably grasp yet this is also very unusual and unnatural compared to the emergent grasp qualities displayed by the policies that have pre-training knowledge on manipulating diverse objects on a generic reposing task objective like in (e-h). Figure 12: Comparison of grasp behaviors. (a, b, c, d) Downstream-task teacher trained from scratch at ADR level 50, with grasps that lack the natural finger gaits emerging from reposing pretraining. (e, f) Pretrained reposing teacher applied zero-shot to FMB peg insertion at ADR level 20, occasionally producing undesirable grasps (e.g., bottom-end, upside-down) that solve the reposing segment but fail at insertion. (g, h) Downstream-task teacher trained with ADEPT at ADR level 50, with natural grasps where the fingers wrap around the peg and caging.