Paper deep dive
Grounding Sim-to-Real Generalization in Dexterous Manipulation: An Empirical Study with Vision-Language-Action Models
Ruixing Jin, Zicheng Zhu, Ruixiang Ouyang, Sheng Xu, Bo Yue, Zhizheng Wu, Guiliang Liu
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 95%
Last extracted: 3/26/2026, 1:40:59 AM
Summary
This paper presents a comprehensive empirical study on Sim-to-Real generalization for Vision-Language-Action (VLA) models in dexterous manipulation. The authors evaluate four key dimensions—domain randomization, photorealistic rendering, physics-realistic modeling, and reinforcement learning fine-tuning—using a unified benchmark and over 10,000 real-world trials. Key findings indicate that spatial feature augmentation is more effective than visual appearance changes, frame-wise randomization outperforms episode-wise strategies, and reinforcement learning significantly enhances robustness to distribution shifts.
Entities (5)
Relation Signals (3)
Reinforcement Fine-Tuning (RFT) → enhances → VLA models
confidence 95% · Reinforcement fine-tuning (RFT) further improves VLA models by optimizing them through interaction with simulated environments
VLA models → evaluatedusing → RoboTwin 2.0
confidence 95% · we build a unified benchmark for VLA systems based on the RoboTwin 2.0 simulation framework
Domain Randomization → improves → Sim-to-Real generalization
confidence 90% · Domain randomization improves robustness to out-of-distribution (OOD) real-world conditions
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Learning a generalist control policy for dexterous manipulation typically relies on large-scale datasets. Given the high cost of real-world data collection, a practical alternative is to generate synthetic data through simulation. However, the resulting synthetic data often exhibits a significant gap from real-world distributions. While many prior studies have proposed algorithms to bridge the Sim-to-Real discrepancy, there remains a lack of principled research that grounds these methods in real-world manipulation tasks, particularly their performance on generalist policies such as Vision-Language-Action (VLA) models. In this study, we empirically examine the primary determinants of Sim-to-Real generalization across four dimensions: multi-level domain randomization, photorealistic rendering, physics-realistic modeling, and reinforcement learning updates. To support this study, we design a comprehensive evaluation protocol to quantify the real-world performance of manipulation tasks. The protocol accounts for key variations in background, lighting, distractors, object types, and spatial features. Through experiments involving over 10k real-world trials, we derive critical insights into Sim-to-Real transfer. To inform and advance future studies, we release both the robotic platforms and the evaluation protocol for public access to facilitate independent verification, thereby establishing a realistic and standardized benchmark for dexterous manipulation policies.
Tags
Links
- Source: https://arxiv.org/abs/2603.22876v1
- Canonical: https://arxiv.org/abs/2603.22876v1
Trouble viewing inline? Open PDF directly →
Full Text
72,914 characters extracted from source content.
Expand or collapse full text
Grounding Sim-to-Real Generalization in Dexterous Manipulation: An Empirical Study with Vision-Language-Action Models Ruixing Jin 1 , Zicheng Zhu 1 , Ruixiang Ouyang 1 , Sheng Xu 1 , Bo Yue 1 , Zhizheng Wu 1 , and Guiliang Liu ⋆1,2 1 School of Data Science, The Chinese University of Hong Kong, Shenzhen 2 Shenzhen Loop Area Institute Abstract. Learning a generalist control policy for dexterous manip- ulation typically relies on large-scale datasets. Given the high cost of real-world data collection, a practical alternative is to generate synthetic data through simulation. However, the resulting synthetic data often ex- hibits a significant gap from real-world distributions. While many prior studies have proposed algorithms to bridge the Sim-to-Real discrepancy, there remains a lack of principled research that grounds these methods in real-world manipulation tasks, particularly their performance on gen- eralist policies such as Vision-Language-Action (VLA) models. In this study, we empirically examine the primary determinants of Sim-to-Real generalization across four dimensions: multi-level domain randomization, photorealistic rendering, physics-realistic modeling, and reinforcement learning updates. To support this study, we design a comprehensive eval- uation protocol to quantify the real-world performance of manipulation tasks. The protocol accounts for key variations in background, light- ing, distractors, object types, and spatial features. Through experiments involving over 10k real-world trials, we derive critical insights into Sim- to-Real transfer. To inform and advance future studies, we release both the robotic platforms and the evaluation protocol for public access to facilitate independent verification, thereby establishing a realistic and standardized benchmark for dexterous manipulation policies. 1 Introduction Developing precise and scalable robotic manipulation policy represents a key milestone toward realizing artificial general intelligence (AGI) [43]. To achieve this goal, Vision-Language-Action (VLA) models have emerged as a principled design for building general-purpose robotic agents, enabling an end-to-end map- ping from visual observations and language instructions to continuous motor actions [25, 61]. VLA methods typically adapt multi-modal foundation mod- els [5,21,50] to embodied control and fine-tune them on diverse robotic demon- strations [16, 39], thereby exhibiting compositional reasoning capabilities from ⋆ Corresponding Author. arXiv:2603.22876v1 [cs.RO] 24 Mar 2026 2F. Author et al. Sim2Real Techniques Domain Randomization 1. Background 2. Distractor 3. Lighting 4. Table Height 5. Camera Pose 6. Randomization Granularity Vision-Language-Action Model Evaluation Sim OOD Experiments 1. Object Initial Position 2. Lighting 3. Table Texture 4. Cross-Object 5. Distractor Do Sim2Real Techniques Truly Close the Gap for VLA Models? Rendering Fidelity 1. Photorealistic 2. Physical-realistic Real-world Experiments Post-Training OpenVLA-OFT ImageLanguage Action Supervised Fine-Tuning RL Fine-Tuning Optional Episode-wise Frame-wise t 1 t 2 t 3 t 1 t 2 t 3 Low Medium High Gravity Static Friction Restitution Dynamic Friction RL vs SFT Episode-wise VS Frame-wise Factored DR Rendering Fidelity Beat Block Place Cup Stack Bowls Pick Dual Bottles Click Bell Fig. 1: Overview of our framework for analyzing Sim2Real generalization in Vision- Language-Action (VLA) models. We study how different Sim2Real techniques, includ- ing domain randomization, rendering fidelity, and reinforcement learning fine-tuning, influence generalization across Vision, Semantics, and Execution under both simulation OOD and real-world evaluations. foundation models [1,18,36,46,63]. This formulation allows a single policy to per- form multi-task manipulation across heterogeneous environments, demonstrating promising language grounding and cross-task generalization [60]. As a learning-based framework for dexterous manipulation, VLA training typically follows a data-driven pipeline. However, this process requires a substan- tial amount of robotic operation trajectories, which are fundamentally scarce, costly to acquire, and inherently hardware-specific. These characteristics make it difficult to scale across diverse tasks, robotic platforms, and real-world sce- narios [6]. To address these challenges, recent studies [9,22,27,59] have explored training VLA models using large-scale simulated data generated by diverse sim- ulation platforms. These platforms include rule-based simulators, such as Mu- JoCo [49] and Isaac Lab [30], as well as learning-based environments built upon world models [45]. Compared to data collection in real-world environments, syn- thetic data can be produced efficiently at scale. However, significant discrepancies in dynamics, kinematics, sensor characteristics, or other factors between simu- lated and real settings result in a pronounced simulation-to-reality (Sim2Real) gap. As a result, VLA models trained in simulation often exhibit degraded per- formance when deployed in real-world scenarios [4,10,34]. Title Suppressed Due to Excessive Length3 The ability to deploy policies trained in simulation to the physical world is therefore essential for scalable and practical robotics. Existing Sim2Real ap- proaches primarily fall into four categories: 1) Domain randomization [7] im- proves robustness to out-of-distribution (OOD) real-world conditions by intro- ducing diverse random perturbations during simulation training. 2) Domain adaptation [2, 13, 58] reduces the discrepancy between simulated and real do- mains by aligning them within a shared feature space. 3) Photo- or physics- realistic rendering mitigates the Sim2Real gap by increasing the visual fidelity of observations and the physical accuracy of environmental dynamics. 4) Re- inforcement fine-tuning (RFT) [19] further improves VLA models by optimiz- ing them through interaction with simulated environments using reinforcement learning (RL). However, these approaches are often evaluated independently on general robotics tasks or control models. It remains unclear which specific factors influence, or to what extent they contribute to, the Sim2Real transferability of VLA models in dexterous manipulation tasks. Such a lack of mechanistic under- standing limits the principled optimization of Sim2Real strategies and hinders systematic diagnosis of VLA model failures under real-world deployment. In this work, we address this question through a factorized empirical anal- ysis of the zero-shot Sim2Real performance of VLA models, where models are trained entirely in simulation and directly deployed to real-world tasks without any real-world data fine-tuning [59]. To evaluate zero-shot Sim2Real trans- fer, we build a unified benchmark for VLA systems based on the RoboTwin 2.0 simulation framework [4]. The benchmark includes a set of representative dual-arm manipulation tasks covering both short-horizon precision interactions and multi-stage behaviors. In addition to simulation OOD evaluation, we con- duct real-world tests under controlled variations, including background textures, lighting perturbations, object instances, distractors, and spatial position changes across predefined grid layouts. Within this unified framework, we conduct a factorized comparison of Sim2Real methods on zero-shot transfer performance, examining the effects of domain ran- domization factors, randomization granularity, photorealistic rendering fidelity, and RL fine-tuning. Through a massive real-world evaluation including more than 10,000 real-world trials, our study offers the following insights: – Spatial features speak louder than appearance. Augmenting spatial features, such as table height and camera pose, yields larger improvements than purely visual perturbations like background textures or lighting. This suggests that spatial variation enhances the policy’s ability to adapt to dif- ferent spatial configurations and strengthens the coupling between visual perception and motor execution. – Granularity matters. Frame-wise domain randomization yields better zero- shot transfer than episode-wise strategies by improving the model’s attention to task-relevant objects rather than background variations. – Simulation fidelity enhances Sim2Real transfer. Our results reinforce the conclusion that higher levels of photorealistic and physics-based fidelity 4F. Author et al. consistently improve Sim2Real performance; however, the magnitude of im- provement diminishes as the simulation fidelity reaches a threshold. – RL enhances robustness to distribution shifts. Compared to SFT- only policies, RFT demonstrates greater resilience to object variations and environmental perturbations. When combined with structured domain ran- domization, the performance gains are further amplified, highlighting the complementary benefits of policy optimization and environmental diversity. Inspired by RoboChallenge [55], we develop an online system that offers open access to standardized protocols, experimental settings, and real-world deploy- ment interfaces of bimanual robots. This platform ensures result reproducibility and enables practitioners to gain deeper insights into Sim2Real transfer. It sup- ports continued investigation and systematic benchmarking of VLA models in physical environments. 2 Related Work 2.1 Vision-Language-Action (VLA) Models Recently, with the advancement of large language models (LLMs) and vision- language models (VLMs) [5, 21, 50], training Vision-Language-Action (VLA) models on large-scale robotic datasets [16,39] has become a significant research direction in the robotics field. Early works such as RT-1 [3] and RT-2 [63] demon- strated the feasibility of training transformer architectures on large-scale robotic datasets by adding action heads to VLMs. Similarly, open-source models such as OpenVLA [18], and CogACT [20] have also proven that VLA architectures can achieve competitive performance across diverse hardware platforms. Meanwhile, recent studies have continuously optimized model architectures; for instance, π 0 [1] adopted flow matching, and OpenVLA-OFT [17] altered the decoding method for action tokens. However, these works primarily focus on simulated environments or deployment fine-tuned on real-world data. Our research will focus on empirical studies of Sim2Real transfer for VLA models. 2.2 Techniques for Sim-to-Real Transfer in Robotics To bridge the gap between simulation and real-world deployment, several main- stream Sim2Real techniques have been widely studied in robotics. Representa- tive approaches include domain randomization, domain adaptation, improving rendering realism, and RL-based post-training. Domain Randomization (DR) remains the foundational approach for zero- shot transfer [11, 48, 62], which expands the simulated distribution by random- izing visual and physical parameters during training to encompass real-world variations. Tobin et al. [48] proposed that randomizing textures, lighting, and camera parameters enables zero-shot transfer for visual grasping tasks by deep neural networks. Subsequently, DR has been widely adopted in object detec- tion [11], robotics [44], autonomous driving [35,41], and other fields. However, DR Title Suppressed Due to Excessive Length5 is typically applied monolithically [4,59,62], where multiple factors are random- ized simultaneously, making it difficult to unravel their individual contributions. Recent studies also have determined relevant parameters by automatic learn- ing [37], active exploration [28], Bayesian update [32, 33], offline inference [47] and continual learning [12] In this work, we address this gap through a factorized empirical study that disentangles and ranks key domain randomization factors to identify their impact on zero-shot Sim2Real transfer for VLA models. Photo/physics realistic Rendering provides a complementary strategy to reduce visual Sim-to-Real discrepancies by improving the realism of simulated observations. Instead of expanding the training distribution as in domain ran- domization, photorealistic simulation aims to narrow the visual reality gap at its source by generating high-fidelity synthetic data. Recent advances in neural scene reconstruction and rendering further enable realistic simulation environ- ments derived from real-world captures [15,51]. In robotics, modern simulation platforms support high-fidelity GPU-accelerated rendering and physics simula- tion [26,52], making photorealism increasingly practical for large-scale training. In this work, we treat rendering fidelity as a controlled experimental factor and systematically evaluate its effect on zero-shot Sim2Real transfer. Reinforcement Fine-Tuning (RFT) has emerged as a powerful paradigm for enhancing the capabilities of foundation models [24]. In large language models (LLMs) and vision-language models (VLMs), RL has been shown to significantly improve reasoning abilities [42] and align outputs with human preferences [14]. Beyond standard supervised fine-tuning (SFT), RL fine-tuning can unlock out- of-distribution generalization and stronger reasoning performance [8,23]. These benefits have recently extended to vision-language-action (VLA) models, where RL fine-tuning has been shown to improve policy performance in both simulation and real-world settings [19, 23]. In this work, we further examine how RL fine- tuning influences zero-shot Sim2Real generalization in VLA models. 2.3 Real-robot Evaluation While simulation platforms are gradually maturing, real-robot evaluation re- mains indispensable. RoboChallenge [54] identifies this as a significant challenge in robotics and has constructed a large-scale online real-robot evaluation in- frastructure, introducing multiple tabletop manipulation tasks. In establishing real-world test sets, Xie et al. [53] discovered that different environmental factors have varying impacts on model performance in the real world, and these factors are largely independent of each other, which means most pairs of factors exhibit no compounding effects. This finding also serves as an important basis for our design of real-world test sets. In this work, we design a controlled benchmark to comprehensively evaluate the zero-shot Sim-to-Real capability of VLA models. 6F. Author et al. 3 Preliminaries 3.1 VLA Models for Robotic Manipulation Robotic manipulation in the Sim2Real setting can be formulated as a Partially Observable Markov Decision Process (POMDP) M = (S,A,O,P T ,R,μ 0 ,γ) where: 1) The state s t ∈S captures the semantic information of a scene, encom- passing the configuration (e.g., layouts, appearance, and physical characteristics) of various types of objects and the robots. 2) The action a t ∈ A corresponds to low-level robot control commands. In our work, actions are parameterized as continuous target joint angles. 3) The observation o t ∈O represents the percep- tual signals captured by sensors. These observations are typically non-Markovian and provide only partial information about the current state. 4) The transition function P T describes the physical dynamics that map the current state s t and executed action a t to the next state s t+1 . 5) The reward function R(s,a) mea- sures task performance and, in many robotic manipulation settings, is defined by task goals (e.g., language instructions) and often simplified as a binary success signal indicating task completion [19]. 6) μ 0 denotes the initial state distribution, and γ ∈ (0, 1] is the discount factor that balances immediate and future rewards. The VLA policy is conditioned on a language instruction l that specifies the task goal and learns a mapping π(a t:t+M | o t ,l) to predict a short-horizon sequence of future actions from the current observation and instruction [17]. The objective is to maximize expected task success under instruction-conditioned trajectories, allowing a unified policy to map high-level semantic commands to low-level motor control. This language interface enables generalization across diverse manipulation goals, object variations, and spatial configurations [1, 17, 18]. 3.2 Zero-Shot Sim2Real Paradigm for VLA Models Let ˆ M = (S,A,O,P,R,μ 0 ,γ) denote the real-world environment, and let M denote the corresponding simulated environment. In simulation, we collect tra- jectories τ = (o 0 ,a 0 ,...,o T ,a T ,l) generated by interacting with M, where l specifies the manipulation task. The policy is optimized in simulation as: π ∗ = arg max π E τ∼M J(τ,π;l) .(1) We present a unified framework for analyzing zero-shot Sim2Real transfer in VLA systems. Our method organizes the problem into two components: policy learning and simulation design. Supervised Fine-Tuning (SFT) We first train the VLA policy in simula- tion using supervised behavioral cloning. Given a demonstration dataset D = (o (i) t ,a (i) t:t+M ,l (i) ), where i denotes the data index, o t denotes the current ob- servation, and l the language instruction, the policy predicts a short-horizon sequence of future continuous actions conditioned on multimodal inputs. Title Suppressed Due to Excessive Length7 Actions are parameterized as continuous low-level control commands (e.g., joint targets and gripper states), forming an action vector a t:t+M ∈ R (M+1)d a over a prediction horizon of M +1 steps. The policy directly outputs continuous action values rather than discretized tokens. The supervised objective minimizes the L1 regression loss [17]: L SFT (θ) = E (o t ,a t:t+M ,l)∼D [∥π θ (o t ,l)− a t:t+M ∥ 1 ],(2) where θ denotes the learnable model parameters. Reinforcement Learning Fine-Tuning To further enhance robustness, we optionally perform reinforcement learning (RL) fine-tuning in simulation. In our study, we adopt Group Relative Policy Optimization (GRPO) [42] as used in SimpleVLA-RL [19]. GRPO optimizes the policy by comparing relative returns within a group of sampled trajectories, encouraging actions that outperform group-level baselines while stabilizing training for large VLA models. Given a task-conditioned reward function R(s t ,l), we optimize the policy us- ing a group-relative objective. For each update, we sample G trajectoriesτ i G i=1 from the current policy and compute trajectory-level returns R i . Instead of us- ing a fixed reference model with KL regularization, we follow the DAPO-style modification and remove the KL penalty to encourage exploration [56]. The policy is optimized via the following clipped objective: J (θ) = E a t ∼π θ old " 1 G G X i=1 1 |τ i | X t min r i,t (θ) ˆ A i , clip r i,t (θ), 1− ε, 1 + ε ˆ A i # , (3) where r i,t (θ) = π θ (a i,t | o i,t ) π θ old (a i,t | o i,t ) , ˆ A i = R i − mean(R j G j=1 ) std(R j G j=1 ) . (4) Here r i,t (θ) denotes the policy likelihood ratio, ˆ A i is the normalized advantage computed from trajectory returns, and ε is a clipping hyperparameter that con- strains r i,t (θ) within [1− ε, 1 + ε] to stabilize policy updates. Only trajectory groups containing both successful and unsuccessful rollouts are used for updates, ensuring meaningful relative comparisons within each group [19]. Notably, Sim2Real transfer in our setting is conducted in a zero-shot man- ner, where no real-world demonstrations are used during training and the learned policy must directly generalize to the real environment ˆ M. However, discrepan- cies between M and ˆ M introduce distribution shifts in visual appearance and geometry. To mitigate this gap, we focus on simulation-side techniques that im- prove robustness without requiring real-world adaptation data. While domain adaptation methods align simulated and real observations using real-world data, they fall outside the zero-shot setting considered in this work. Instead, we adopt two complementary strategies for constructing simulation training data: Do- main Randomization, which exposes the policy to diverse environment con- figurations, and Rendering Fidelity, which improves the realism of simulated observations, including photorealism and physical realism. 8F. Author et al. Domain Randomization (DR). We train policies over a distribution of simu- lator parameters rather than a fixed environment [6]. Let x denote an observation generated by the simulator, and let ξ represent the simulator parameters control- ling the scene configuration. In a fixed simulator, observations follow x∼ p ξ (x). Under DR, simulator parameters are sampled from a distribution ξ ∼ p(ξ), leading to a training distribution x∼ R p(x| ξ)p(ξ)dξ. The learning objective for a policy f φ becomes L = E ξ∼p(ξ) E x∼p(x|ξ) ℓ(f φ (x),y) . (5) The objective of DR is not to exactly match the real distribution p real (x), but to expand the synthetic distribution such that supp(p real )⊆ supp(p DR ), so that real-world observations appear as a special case within the randomized synthetic distribution. Following structured domain randomization principles from prior manufac- turing DR literature [11,62], we factorize domain randomization into structured components as ξ =ξ bg ,ξ dist ,ξ cam ,ξ light ,ξ table . Domain Randomization Factors. We decompose domain randomization into five structured components to analyze their individual contributions: Background (ξ BG ), which randomizes wall and table textures from a predefined image set; Table Distractor (ξ TD ), which samples the number, mesh, pose, and texture of irrelevant objects while avoiding overlap with task-relevant ones; Camera Pose (ξ CP ), which applies random translational offsets to the head camera position at initialization; Lighting (ξ LT ), which randomizes light color and position within scene bounds; and Table Height (ξ TH ), which samples table height from a predefined range. Randomization Granularity. We further study the temporal scope of random- ization by comparing two strategies: episode-wise, where factors are sampled once per episode and fixed during rollout, and frame-wise, where factors are resampled at every simulation step. Rendering Fidelity. Beyond domain randomization, the fidelity of both visual rendering and physics simulation plays a critical role in bridging the Sim2Real gap. Photorealism influences the statistical properties of simulated observa- tions by improving light transport, shadow consistency, and global illumination, producing images that more closely resemble real-world sensor data [15]. In prac- tice, we vary rendering configurations—including ray tracing (RT), samples per pixel, path depth, and denoising—to control the level of visual realism. Physical realism affects the accuracy of object dynamics and contact interactions during manipulation. We vary key physics parameters such as gravitational accelera- tion, static/dynamic friction coefficients, and restitution, which govern object motion and contact behavior in simulation. Together, these factors determine how closely the simulated environment approximates real-world conditions. Title Suppressed Due to Excessive Length9 3.3 Problem Formulation: Empirical Study for Sim2Real VLAs Real-world robotic data is scarce, expensive to collect, and highly environment- specific [6], making large-scale training of Vision-Language-Action (VLA) mod- els using real demonstrations impractical. As a result, simulation becomes the primary scalable source of training data. However, discrepancies between simulated and real environments often cause policies trained in simulation to degrade when deployed in the physical world [4, 10, 34]. Although some sim-to-real techniques, such as domain randomization, have shown promising results in robotics [29,31,59], its role in large-scale VLA models remains insufficiently understood. Compared with conventional control policies, VLA models jointly integrate visual perception, language grounding, and long-horizon decision making, which may introduce new sensitivities to sim- ulation–reality discrepancies. This motivates a systematic empirical study to disentangle how individual simulation design factors affect Sim2Real transfer in VLA models. 4 Experiments 4.1 Evaluation Protocol. We conduct a systematic study of zero-shot simulation-to-real generalization under controlled distribution shifts. All experiments use the OpenVLA-OFT policy [17] trained with supervised fine-tuning (SFT) in simulation. Within this framework, we evaluate how structured domain randomization, randomization granularity, and rendering fidelity affect Sim2Real transfer in both simulation and real-world environments. Simulation Dataset. All policies are trained in RobotWin 2.0 [4], a physics- based manipulation simulator that supports configurable lighting, camera pose, table height, background textures, and distractor objects. We evaluate five manipulation tasks with varying horizon lengths and spatial complexity. For each task and each training setting, we collect 100 demonstration trajectories in simulation. The simulated robot embodiment uses the Cobot Magic platform to ensure consistency with real-world deployment [4]. Visual input is provided by a sin- gle forward-facing Intel RealSense D435 RGB camera at 640×480 resolution, adopted for simplicity and clearer interpretation of experimental results. Camera intrinsics remain fixed across experiments, while camera pose may be random- ized depending on the experimental condition. Policy Architecture We base our study on OpenVLA-OFT [17], a state-of- the-art Vision-Language-Action model that enhances the original OpenVLA [18] with optimized designs for efficient and robust control. Built on a fused visual encoder and a Llama2 7B language backbone [50], OpenVLA-OFT incorporates 10F. Author et al. Manipulation Objects Distractors Lighting equipment Camera Tablecloth Robot Object Initial Position Lighting Background Direction: 1 Direction: 2Direction: 3Color Change White(canonical) Object Distractor Level 1 Pos: 1 Pos: 2 Pos: 3 Pos: 4 Wood Table Yellow GridBlue Seen Unseen :1Unseen :2 Unseen :3 Level 2 Level 3 Different set Fig. 2: Real-world manipulation setup and randomized factors used in our experiments. Left: the physical platform with robot arms, camera, lighting, objects, and distractors. Right: examples of variations including object positions, lighting, backgrounds, object instances, and distractor configurations. two key improvements: (1)Parallel decoding with action chunking, which replaces autoregressive generation to enable single-pass prediction of multiple future ac- tions. (2) Continuous action representation with L1 regression, which avoids information loss from discretization and enables finer-grained control. Simulation OOD Evaluation. To measure robustness within simulation, we construct Out-of-Distribution (Sim-OOD) environments by introducing unseen variations along individual factors. These include unseen background and ta- ble textures, unseen distractor objects, table height perturbations, camera pose shifts, and lighting disturbances. Each factor is varied independently to isolate its effect on performance. Sim OOD evaluation serves as a controlled proxy for real-world generalization. Real-World Evaluation Setup. Zero-shot real-world evaluation is conducted on a physical Piper robot platform equipped with a single RealSense D435 RGB camera (640×480 input resolution). An overview of the real-world setup is shown in Fig. 2. The workspace consists of an adjustable-height table with replaceable tablecloth textures and configurable lighting. We evaluate robustness across multiple real-world factors: – Background variation: three table textures (wood, blue fabric, yellow grid). – Lighting variation: three distinct light positions with varying illumination colors. – Object variation: both seen and unseen object instances within the same task category, evaluating generalization to novel appearances. Title Suppressed Due to Excessive Length11 Table 1: Factorized domain randomization results under Simulation OOD and zero- shot real-world evaluation across five tasks. Click BellPlace Empty CupBeat Block HammerStack Bowls TwoPlace Dual Bottles Data settingSim-OOD Real Sim-OOD RealSim-OOD RealSim-OOD RealSim-OOD Real Clean14% 2.7%11%5.4%17%0.0%36% 26.2%19%1.5% BG23% 11.5%15% 10.2%30%2.7%42% 40.4%27%12.3% LT22% 12.3% 12%8.6%29%4.6%43% 31.5%24%7.3% TD15% 3.1%12%6.9%23%0.0%38% 27.3%21%4.6% CP34% 23.5%20% 17.5%32%4.6%47% 42.3%30% 16.2% TH40% 36.9% 26% 24.8%37%6.5%58% 49.6%32% 15.4% TH + CP + BG 50% 47.7%34% 33.5%43%8.5%62% 60.0%42% 21.2% TH + CP + LT 48% 44.2%31% 27.9%46%7.3%60% 54.6%39%20.0% TH + CP + TD 43% 40.0%29% 26.3%38%6.5%56% 52.7%32%16.5% All Factors54% 49.7%44% 41.0%49% 11.5%65% 63.1%52% 23.8% – Distractor presence: task-irrelevant objects placed in the scene at three difficulty levels (2, 4, and 8 distractors), sampled from a separate object set distinct from task objects. – Spatial generalization: evaluation on unseen positions within a predefined 3× 3 grid. For each task, we first evaluate the policy under a canonical base configura- tion. We then systematically vary one factor at a time while keeping all other factors fixed. For each configuration, we evaluate the task under multiple object positions by applying small random shifts, and report the resulting success rate across these variations. 4.2 Domain Randomization for Sim2Real Generalization Effects of Domain Randomization To analyze the role of domain random- ization, we decompose the five randomization factors into two categories: appear- ance factors, including Background (BG), Lighting (LT), and Table Distractor (TD); and spatial factors, including Table Height (TH) and Camera Position (CP). In this experiment, all randomization is applied in an episode-wise man- ner, where parameters are sampled once at scene initialization and kept fixed throughout each rollout. Tab. 1 reports the success rates under Simulation OOD and zero-shot real- world evaluation across different domain randomization factors. Training with clean data leads to poor generalization, resulting in low success rates in both Sim-OOD and real environments. Introducing individual randomization factors consistently improves performance across tasks. Lesson 1: Spatial randomization is the primary driver of Sim2Real generalization. Among all factors, spatial perturbations such as camera pose (CP) and table height (TH) provide the largest performance gains, significantly improving success rates across both Sim-OOD and real-world evaluations. In contrast, appearance-level changes such as background (BG) and lighting (LT) 12F. Author et al. Table 2: Results of randomization granularity.∆ denotes the gain of frame-wise over episode-wise. Click BellPlace Empty CupBeat Block HammerStack Bowls TwoPick Dual Bottles Data SettingReal Sim-OODReal Sim-OODReal Sim-OODReal Sim-OODReal Sim-OOD Episode-wise CP23.5% 34%17.5% 20%4.6%32%42.3% 47%16.2% 30% Frame-wise CP31.9% 37%25.6% 27%7.3%34%49.2% 51%19.6% 34% ∆ (Frame− Episode)8.4%3.0%8.1%7.0%2.7%2.0%6.9%4.0%3.4%4.0% Episode-wise BG11.5% 23%10.2% 15%2.7%30%40.4% 42%12.3% 27% Frame-wise BG15.0% 31% 23.1% 27%7.7%41%45.0% 53%14.2% 32% ∆ (Frame− Episode)3.5%8.0%12.9%12.0%5.0%11.0%4.6%11.0%1.9%5.0% Episode-wise LT12.3% 22%8.6%12%4.6%29%31.5% 43%7.3%24% Frame-wise LT14.6% 23%9.0%13%5.0%30%32.6% 43%7.3%26% ∆ (Frame− Episode)2.3%1.0%0.4%1.0%0.4%1.0%1.1%0.0%0.0%2.0% yield smaller improvements when applied individually. This suggests that varia- tions affecting geometric relationships and viewpoint are more critical for robust policy learning than purely visual changes. Lesson 2: Appearance randomization provides complementary ben- efits when combined with spatial perturbations. Although background and lighting perturbations alone produce relatively modest gains, combining them with spatial randomization further improves performance. In particular, configurations such as TH + CP + BG and TH + CP + LT consistently out- perform single-factor settings across most tasks. The best results are achieved when all factors are applied simultaneously, yielding the highest success rates in both Sim-OOD and real-world evaluations. These results indicate that while spatial variations play a dominant role in improving policy robustness, appearance-level randomization helps cover addi- tional visual discrepancies between simulation and reality when combined with spatial perturbations. Effects of Randomization Granularity Building upon the episode-wise set- ting used in Experiment 1, we next investigate whether increasing the temporal frequency of randomization further improves zero-shot Sim2Real transfer. Specif- ically, we consider episode-wise and frame-wise strategies as two representative boundary cases of temporal stochasticity. In this study, we limit frame-wise ran- domization to background textures, lighting configuration, and camera position. Factors that directly affect task feasibility, including distractor placement and table height, remain episode-wise to preserve task stability, maintain consistent task planning, and avoid disrupting motion execution during rollout. Lesson 3: Granularity matters. Applying randomization at the frame level consistently outperforms episode-wise randomization across tasks ( Tab. 2). Frame-wise perturbations introduce continuous variation within each rollout, preventing the policy from overfitting to static scene configurations. This ef- fect is particularly evident for background (BG) and camera pose (CP), where frame-wise randomization improves real-world success rates by roughly 3–8% for CP and up to around 13% for BG, with similarly consistent gains observed Title Suppressed Due to Excessive Length13 under Sim-OOD evaluation. In contrast, lighting perturbations provide compara- tively smaller improvements, typically around 0–2%, suggesting that illumination changes alone contribute less to generalization when applied at a finer temporal granularity. 4.3 Rendering Fidelity for Sim2Real Generalization We next investigate whether increasing rendering fidelity improves zero-shot Sim2Real transfer ( Fig. 3b). We compare three rendering presets that differ in ray-tracing activation and sampling configuration: The Low setting disables RT and uses low sampling and path depth. The Medium setting enables RT with the same sampling and path depth as the Low setting, while introducing additional rendering effects. The High setting further increases the sampling rate and path depth to achieve higher visual fidelity. In addition to visual fidelity, we also consider physical realism. To examine the effect of physics fidelity, we construct a setting with intentionally distorted dynamics by modifying key physical parameters, including gravitational acceler- ation, static and dynamic friction coefficients, and restitution. These parameters are set to relatively extreme values to reduce the physical realism of the sim- ulator, allowing us to evaluate how violations of realistic dynamics affect VLA training and Sim2Real transfer. Lesson 4: Simulation fidelity enhances Sim2Real transfer. Fig. 3a reports the performance comparison across five tasks under different rendering fidelity settings. Increasing photorealism consistently improves real-world success rates, indicating that more realistic visual appearance helps the policy learn representations that transfer better to the physical environment. However, the improvement becomes marginal beyond the medium setting, suggesting that moderate photorealism already captures most transferable visual cues. The figure also shows that improving physical realism leads to certain gains in Sim2Real performance for VLA models. However, the improvement is smaller compared with that obtained from increasing photorealism, indicating that vi- sual fidelity plays a more dominant role in facilitating Sim2Real transfer. 4.4 Reinforcement Learning for Sim2Real Generalization Finally, we evaluate the effect of reinforcement learning (RL) fine-tuning on zero-shot Sim2Real transfer. Following the design in SimpleVLA [19], we modify the action prediction head of OpenVLA-OFT by replacing the original MLP- based continuous action regression with the LLaMA2 output head that generates discretized action tokens optimized with a cross-entropy loss. Based on this modification, we first re-train an SFT policy using 100 demon- strations per task, which serves as the initialization for all RL experiments. Starting from this SFT checkpoint, we further perform RL fine-tuning in simu- lation under different rollout settings. We compare three training variants: (1) 14F. Author et al. (a) Low photorealism Medium photorealism High photorealism (b) Fig. 3: Rendering fidelity analysis. (a) Quantitative results showing the impact of pho- torealism and physical realism on Sim-OOD and real-world success rates. (b) Example renderings under different photorealism levels (Low, Medium, High). Table 3: Effect of reinforcement learning (RL) and domain randomization (DR) on zero-shot Sim2Real performance across five tasks. Click BellPlace Empty CupBeat Block HammerStack Bowls TwoPlace Dual Bottles Training Variant Sim-OOD RealSim-OOD RealSim-OOD RealSim-OOD RealSim-OOD Real SFT11%0%23%6.8%15%0%33% 21.2%10%0% SFT + RL42% 36.9%65% 34.8%52%9.6%60% 59.2%48%26.5% SFT + RL + DR 60% 50.8%87% 51.4%72% 16.2%68% 64.6%67% 30.8% SFT trained in a clean simulation environment, (2) SFT followed by RL fine- tuning using rollouts collected in a clean simulation environment, and (3) SFT followed by RL fine-tuning using rollouts collected under domain randomization. Lesson 5: RL enhances robustness to distribution shifts. As shown in Tab. 3, even when RL is trained under clean simulation conditions, it sub- stantially improves performance over the SFT baseline, increasing the average real-world success rate from 5.6% to 33.4%. Notably, this clean RL setting al- ready approaches the performance achieved by SFT combined with domain ran- domization as shown in Tab. 1, suggesting that RL itself introduces stronger policy generalization beyond supervised imitation. Furthermore, when domain randomization is incorporated during RL rollouts, the performance further im- proves to 42.8% in real-world evaluation and 70.8% in Sim-OOD, indicating that RL can effectively exploit diverse training experiences to further enhance Sim2Real robustness. 5 Conclusion In this work, we systematically study how different Sim2Real techniques affect the generalization of Vision-Language-Action (VLA) models. Our experiments show that spatial domain randomization plays a dominant role in improving ro- bustness, while appearance randomization provides complementary benefits. We further find that frame-wise randomization is more effective than episode-wise Title Suppressed Due to Excessive Length15 perturbations, higher simulation fidelity enhances Sim2Real transfer, and rein- forcement learning fine-tuning significantly improves robustness to distribution shifts. In addition, we develop an online benchmarking platform that provides stan- dardized protocols, experimental settings, and real-world deployment interfaces for bimanual robots, enabling reproducible evaluation and continued benchmark- ing of Sim2Real performance for VLA models. References 1. Black, K., Brown, N., Driess, D., Esmail, A., Equi, M., Finn, C., Fusai, N., Groom, L., Hausman, K., Ichter, B., Jakubczak, S., Jones, T., Ke, L., Levine, S., Li-Bell, A., Mothukuri, M., Nair, S., Pertsch, K., Shi, L.X., Tanner, J., Vuong, Q., Walling, A., Wang, H., Zhilinsky, U.:π0: A vision-language-action flow model for general robot control. ArXiv abs/2410.24164 (2024) 2. Bousmalis, K., Irpan, A., Wohlhart, P., Bai, Y., Kelcey, M., Kalakrishnan, M., Downs, L., Ibarz, J., Pastor, P., Konolige, K., Levine, S., Vanhoucke, V.: Using simulation and domain adaptation to improve efficiency of deep robotic grasping. In: International Conference on Robotics and Automation, ICRA. p. 4243–4250 (2018) 3. Brohan, A., Brown, N., Carbajal, J., Chebotar, Y., Dabis, J., Finn, C., Gopalakr- ishnan, K., Hausman, K., Herzog, A., Hsu, J., Ibarz, J., Ichter, B., Irpan, A., Jack- son, T., Jesmonth, S., Joshi, N.J., et al.: Rt-1: Robotics transformer for real-world control at scale. ArXiv abs/2212.06817 (2022) 4. Chen, T., Chen, Z., Chen, B., Cai, Z., Liu, Y., Liang, Q., Li, Z., Lin, X., Ge, Y., Gu, Z., Deng, W., Guo, Y., Nian, T., Xie, X., Chen, Q., Su, K., Xu, T., Liu, G., Hu, M., ang Gao, H., Wang, K., Liang, Z., Qin, Y., Yang, X., Luo, P., Mu, Y.: Robotwin 2.0: A scalable data generator and benchmark with strong domain randomization for robust bimanual robotic manipulation. ArXiv abs/2506.18088 (2025) 5. Chen, X., Djolonga, J., Padlewski, P., Mustafa, B., Changpinyo, S., Wu, J., Ruiz, C.R., Goodman, S., Wang, X., Tay, Y., Shakeri, S., Dehghani, M., otherss: Pali-x: On scaling up a multilingual vision and language model. ArXiv abs/2305.18565 (2023) 6. Chen, X., Hu, J., Jin, C., Li, L., Wang, L.: Understanding domain randomization for sim-to-real transfer. In: International Conference on Learning Representations, ICLR (2022) 7. Chen, X., Hu, J., Jin, C., Li, L., Wang, L.: Understanding domain randomization for sim-to-real transfer. In: International Conference on Learning Representations, ICLR (2022) 8. Chu, T., Zhai, Y., Yang, J., Tong, S., Xie, S., Schuurmans, D., Le, Q.V., Levine, S., Ma, Y.: SFT memorizes, RL generalizes: A comparative study of foundation model post-training. In: International Conference on Machine Learning, ICML. vol. 267 (2025) 9. Deng, S., Yan, M., Wei, S., Ma, H., Yang, Y., Chen, J., Zhang, Z., Yang, T., Zhang, X., Zhang, W., et al.: Graspvla: a grasping foundation model pre-trained on billion-scale synthetic action data. arXiv preprint arXiv:2505.03233 (2025) 10. ud Din, M., Akram, W., Saoud, L.S., Rosell, J., Hussain, I.: Vision language action models in robotic manipulation: A systematic review. ArXiv abs/2507.10672 (2025) 16F. Author et al. 11. Fan, J., Xie, Y., Li, Z.: Domain randomization for object detection: A survey. In: International Conference on Advanced Electronic Materials, Computers and Software Engineering , AEMCSE (2025) 12. Josifovski, J., Auddy, S., Malmir, M., Piater, J.H., Knoll, A., Navarro-Guerrero, N.: Continual domain randomization. In: IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS. p. 4965–4972 (2024) 13. Josifovski, J., Gu, S., Malmir, M., Huang, H., Auddy, S., Navarro-Guerrero, N., Spanos, C., Knoll, A.: Safe continual domain adaptation after sim2real transfer of reinforcement learning policies in robotics. arXiv preprint arXiv:2503.10949 (2025) 14. Kaufmann, T., Weng, P., Bengs, V., Hüllermeier, E.: A survey of reinforcement learning from human feedback. ArXiv abs/2312.14925 (2023) 15. Kerbl, B., Kopanas, G., Leimkühler, T., Drettakis, G., et al.: 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph. (2023) 16. Khazatsky, A., et al.: Droid: A large-scale in-the-wild robot manipulation dataset. ArXiv abs/2403.12945 (2024), https://api.semanticscholar.org/CorpusID: 268531351 17. Kim, M.J., Finn, C., Liang, P.: Fine-tuning vision-language-action models: Opti- mizing speed and success. ArXiv abs/2502.19645 (2025) 18. Kim, M.J., Pertsch, K., Karamcheti, S., Xiao, T., Balakrishna, A., Nair, S., Rafailov, R., Foster, E.P., Lam, G., Sanketi, P.R., Vuong, Q., Kollar, T., Burchfiel, B., Tedrake, R., Sadigh, D., Levine, S., Liang, P., Finn, C.: Openvla: An open- source vision-language-action model. ArXiv abs/2406.09246 (2024) 19. Li, H., Zuo, Y., Yu, J., Zhang, Y., Yang, Z., Zhang, K., Zhu, X., Zhang, Y., Chen, T., Cui, G., Wang, D., Luo, D., Fan, Y., Sun, Y., Zeng, J., Pang, J., Zhang, S., Wang, Y., Mu, Y., Zhou, B., Ding, N.: Simplevla-rl: Scaling vla training via reinforcement learning. In: International Conference on Learning Representations, ICLR (2026) 20. Li, Q., Liang, Y., Wang, Z., Luo, L., Chen, X., Liao, M., Wei, F., Deng, Y., Xu, S., Zhang, Y., et al.: Cogact: A foundational vision-language-action model for synergiz- ing cognition and action in robotic manipulation. arXiv preprint arXiv:2411.19650 (2024) 21. Li, Z., Chen, G., Liu, S., Wang, S., Vibashan, V., Ji, Y., Lan, S., Zhang, H., Zhao, Y., Radhakrishnan, S., Chang, N., Sapra, K., Deshmukh, A.S., Rintamaki, T., Le, M., Karmanov, I., Voegtle, L., Fischer, P., Huang, D.A., Roman, T., Lu, T., Álvarez, J.M., Catanzaro, B., Kautz, J., Tao, A., Liu, G., Yu, Z.: Eagle 2: Build- ing post-training data strategies from scratch for frontier vision-language models. ArXiv abs/2501.14818 (2025), https://api.semanticscholar.org/CorpusID: 275921957 22. Liu, G., Deng, Y., Zhao, R., Zhou, H., Chen, J., Chen, J., Xu, R., Tai, Y., Jia, K.: Dexscale: Automating data scaling for sim2real generalizable robot control. In: International Conference on Machine Learning, ICML (2025) 23. Liu, J., Gao, F., Wei, B., Chen, X., Liao, Q., Wu, Y., Yu, C., Wang, Y.: What can RL bring to VLA generalization? an empirical study. In: Conference and Workshop on Neural Information Processing Systems, NeurIPS (2025) 24. Liu, K., Yang, D., Qian, Z., Yin, W., Wang, Y., Li, H., Liu, J., Zhai, P., Liu, Y., Zhang, L.: Reinforcement learning meets large language models: A survey of advancements and applications across the llm lifecycle. ArXiv abs/2509.16679 (2025), https://api.semanticscholar.org/CorpusID:281421479 25. Ma, Y., Song, Z., Zhuang, Y., Hao, J., King, I.: A survey on vision-language-action models for embodied ai. arXiv preprint arXiv:2405.14093 (2024) Title Suppressed Due to Excessive Length17 26. Makoviychuk, V., Wawrzyniak, L., Guo, Y., Lu, M., Storey, K., Macklin, M., Hoeller, D., Rudin, N., Allshire, A., Handa, A., State, G.: Isaac gym: High perfor- mance GPU based physics simulation for robot learning. In: Thirty-fifth Confer- ence on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2) (2021) 27. Mandlekar, A., Nasiriany, S., Wen, B., Akinola, I., Narang, Y., Fan, L., Zhu, Y., Fox, D.: Mimicgen: A data generation system for scalable robot learning using human demonstrations. In: Annual Conference on Robot Learning, CoRL (2023) 28. Mehta, B., Diaz, M., Golemo, F., Pal, C.J., Paull, L.: Active domain randomization. In: Conference on Robot Learning, CoRL. vol. 100, p. 1162–1176 (2019) 29. Mehta, B., Diaz, M., Golemo, F., Pal, C.J., Paull, L.: Active domain randomization. ArXiv abs/1904.04762 (2019), https://api.semanticscholar.org/CorpusID: 104291994 30. Mittal, M., Roth, P., Tigue, J., Richard, A., Zhang, O., Du, P., Serrano-Muñoz, A., Yao, X., Zurbrüg, R., Rudin, N., Wawrzyniak, L., Rakhsha, M., Denzler, A., Heiden, E., Borovicka, A., Ahmed, O., Akinola, I., Anwar, A., Carlson, M.T., Feng, J.Y., Garg, A., Gasoto, R., Gulich, L., Guo, Y., Gussert, M., Hansen, A., Kulkarni, M., Li, C., Liu, W., Makoviychuk, V., Malczyk, G., Mazhar, H., Moghani, M., Murali, A., Noseworthy, M., Poddubny, A., Ratliff, N., Rehberg, W., Schwarke, C., Singh, R., Smith, J.L., Tang, B., Thaker, R., Trepte, M., Wyk, K.V., Yu, F., Millane, A., Ramasamy, V., Steiner, R., Subramanian, S., Volk, C., Chen, C., Jawale, N., Kuruttukulam, A.V., Lin, M.A., Mandlekar, A., Patzwaldt, K., Welsh, J., Zhao, H., Anes, F., Lafleche, J.F., Moënne-Loccoz, N., Park, S., Stepinski, R., Gelder, D.V., Amevor, C., Carius, J., Chang, J., Chen, A.H., de Heras Ciechomski, P., Daviet, G., Mohajerani, M., von Muralt, J., Reutskyy, V., Sauter, M., Schirm, S., Shi, E.L., Terdiman, P., Vilella, K., Widmer, T., Yeoman, G., Chen, T., Grizan, S., Li, C., Li, L., Smith, C., Wiltz, R., Alexis, K., Chang, Y., Chu, D., Fan, L.J., Farshidian, F., Handa, A., Huang, S., Hutter, M., Narang, Y., Pouya, S., Sheng, S., Zhu, Y., Macklin, M., Moravanszky, A., Reist, P., Guo, Y., Hoeller, D., State, G.: Isaac lab: A gpu-accelerated simulation framework for multi-modal robot learning. arXiv preprint arXiv:2511.04831 (2025), https://arxiv.org/abs/2511.04831 31. Muratore, F., Eilers, C., Gienger, M., Peters, J.: Data-efficient domain random- ization with bayesian optimization. IEEE Robotics and Automation Letters 6, 911–918 (2020), https://api.semanticscholar.org/CorpusID:230524079 32. Muratore, F., Eilers, C., Gienger, M., Peters, J.: Data-efficient domain random- ization with bayesian optimization. IEEE Robotics and Automation Letters 6(2), 911–918 (2021) 33. Muratore, F., Gruner, T., Wiese, F., Belousov, B., Gienger, M., Peters, J.: Neu- ral posterior domain randomization. In: Conference on Robot Learning, CoRL. vol. 164, p. 1532–1542 (2021) 34. Nasiriany, S., Maddukuri, A., Zhang, L., Parikh, A., Lo, A., Joshi, A., Man- dlekar, A., Zhu, Y.: Robocasa: Large-scale simulation of everyday tasks for gen- eralist robots. ArXiv abs/2406.02523 (2024), https://api.semanticscholar. org/CorpusID:270226600 35. Niu, H., Hu, J., Cui, Z., Zhang, Y.: Dr2l: Surfacing corner cases to robustify au- tonomous driving via domain randomization reinforcement learning. In: Interna- tional Conference on Computer Science and Application Engineering (2021) 36. Nvidia, Bjorck, J., Castañeda, F., Cherniadev, N., Da, X., Ding, R., LinxiJimFan, Fang, Y., Fox, D., Hu, F., Huang, S., Jang, J., Jiang, Z., Kautz, J., Kundalia, K., Lao, L., Li, Z., Lin, Z., Lin, K., Liu, G., Llontop, E., Magne, L., Mandlekar, 18F. Author et al. A., Narayan, A., Nasiriany, S., Reed, S., Tan, Y.L., Wang, G., Wang, Z., Wang, J., Wang, Q., Xiang, J., Xie, Y., Xu, Y., Xu, Z.T., Ye, S., Yu, Z., Zhang, A., Zhang, H., Zhao, Y., Zheng, R., Zhu, Y.: Gr00t n1: An open foundation model for generalist humanoid robots. ArXiv abs/2503.14734 (2025), https://api. semanticscholar.org/CorpusID:277113335 37. OpenAI, Akkaya, I., Andrychowicz, M., Chociej, M., Litwin, M., McGrew, B., Petron, A., Paino, A., Plappert, M., Powell, G., Ribas, R., Schneider, J., Tezak, N., Tworek, J., Welinder, P., Weng, L., Yuan, Q., Zaremba, W., Zhang, L.: Solving rubik’s cube with a robot hand. CoRR abs/1910.07113 (2019) 38. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023) 39. Padalkar, A., et al.: Open x-embodiment: Robotic learning datasets and rt-x mod- els : Open x-embodiment collaboration0. International Conference on Robotics and Automation ,ICRA p. 6892–6903 (2023), https://api.semanticscholar.org/ CorpusID:263626099 40. Perez, E., Strub, F., De Vries, H., Dumoulin, V., Courville, A.: Film: Visual rea- soning with a general conditioning layer. In: Proceedings of the AAAI conference on artificial intelligence. vol. 32 (2018) 41. Pouyanfar, S., Saleem, M., George, N., Chen, S.C.: Roads: Randomization for obstacle avoidance and driving in simulation. In: Conference on Computer Vision and Pattern Recognition Workshops, CVPRW (2019) 42. Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J.M., Zhang, M., Li, Y.K., Wu, Y., Guo, D.: Deepseekmath: Pushing the limits of mathematical reasoning in open language models. ArXiv abs/2402.03300 (2024) 43. Smith, C., Karayiannidis, Y., Nalpantidis, L., Gratal, X., Qi, P., Dimarogonas, D.V., Kragic, D.: Dual arm manipulation - A survey. Robotics and Autonomous Systems 60(10), 1340–1353 (2012) 44. Tan, J., Zhang, T., Coumans, E., Iscen, A., Bai, Y., Hafner, D., Bohez, S., Van- houcke, V.: Sim-to-real: Learning agile locomotion for quadruped robots. ArXiv abs/1804.10332 (2018) 45. Team, G., Ye, A., Wang, B., Ni, C., Huang, G., Zhao, G., Li, H., Zhu, J., Li, K., Xu, M., et al.: Gigaworld-0: World models as data engine to empower embodied ai. arXiv preprint arXiv:2511.19861 (2025) 46. Team, O.M., Ghosh, D., Walke, H.R., Pertsch, K., Black, K., Mees, O., Dasari, S., Hejna, J., Kreiman, T., Xu, C., Luo, J., Tan, Y.L., Sanketi, P.R., Vuong, Q., Xiao, T., Sadigh, D., Finn, C., Levine, S.: Octo: An open-source generalist robot policy. ArXiv abs/2405.12213 (2024), https://api.semanticscholar.org/CorpusID: 266379116 47. Tiboni, G., Arndt, K., Kyrki, V.: DROPO: sim-to-real transfer with offline domain randomization. Robotics and Autonomous Systems 166, 104432 (2023) 48. Tobin, J., Fong, R., Ray, A., Schneider, J., Zaremba, W., Abbeel, P.: Domain ran- domization for transferring deep neural networks from simulation to the real world. In: International Conference on Intelligent Robots and Systems, IROS (2017) 49. Todorov, E., Erez, T., Tassa, Y.: Mujoco: A physics engine for model-based control. In: IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS. p. 5026–5033 (2012) 50. Touvron, H., et al.: Llama 2: Open foundation and fine-tuned chat models. ArXiv abs/2307.09288 (2023), https://api.semanticscholar.org/CorpusID: 259950998 Title Suppressed Due to Excessive Length19 51. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4d gaussian splatting for real-time dynamic scene rendering. In: conference on computer vision and pattern recognition, CVPR (2024) 52. Xiang, F., Qin, Y., Mo, K., Xia, Y., Zhu, H., Liu, F., Liu, M., Jiang, H., Yuan, Y., Wang, H., et al.: Sapien: A simulated part-based interactive environment. In: conference on computer vision and pattern recognition, CVPR (2020) 53. Xie, A., Lee, L., Xiao, T., Finn, C.: Decomposing the generalization gap in imi- tation learning for visual robotic manipulation. Conference on Robotics and Au- tomation, ICRA (2023) 54. Yakefu, A., Xie, B., Xu, C., Zhang, E., Zhou, E., Jia, F., Yang, H., Fan, H., Zhang, H., Peng, H., Tan, J., Huang, J., et al.: Robochallenge: Large-scale real-robot evaluation of embodied policies. ArXiv abs/2510.17950 (2025) 55. Yakefu, A., Xie, B., Xu, C., Zhang, E., Zhou, E., Jia, F., Yang, H., Fan, H., Zhang, H., Peng, H., et al.: Robochallenge: Large-scale real-robot evaluation of embodied policies. arXiv preprint arXiv:2510.17950 (2025) 56. Yu, Q., Zhang, Z., et al.: Dapo: An open-source llm reinforcement learning system at scale. ArXiv abs/2503.14476 (2025), https://api.semanticscholar.org/ CorpusID:277104124 57. Zhai, X., Mustafa, B., Kolesnikov, A., Beyer, L.: Sigmoid loss for language im- age pre-training. In: Proceedings of the IEEE/CVF international conference on computer vision. p. 11975–11986 (2023) 58. Zhang, J., Tai, L., Yun, P., Xiong, Y., Liu, M., Boedecker, J., Burgard, W.: Vr-goggles for robots: Real-to-sim domain adaptation for visual control. IEEE Robotics and Automation Letters 4(2), 1148–1155 (2019) 59. Zhao, R., Xu, S., Jin, R., Deng, Y., Tai, Y., Jia, K., Liu, G.: Sim2real VLA: Zero- shot generalization of synthesized skills to realistic manipulation. In: International Conference on Learning Representations, ICLR (2026) 60. Zhao, W.X., Zhou, K., Li, J., Tang, T., Wang, X., Hou, Y., Min, Y., Zhang, B., Zhang, J., Dong, Z., Du, Y., Yang, C., Chen, Y., Chen, Z., Jiang, J., Ren, R., Li, Y., Tang, X., Liu, Z., Liu, P., Nie, J., rong Wen, J.: A survey of large language models. ArXiv abs/2303.18223 (2023), https://api.semanticscholar.org/CorpusID: 257900969 61. Zheng, Y., Yao, L., Su, Y., Zhang, Y., Wang, Y., Zhao, S., Zhang, Y., Chau, L.P.: A survey of embodied learning for object-centric robotic manipulation. Machine Intelligence Research p. 1–39 (2025) 62. Zhu, X., Henningsson, J., Li, D., Mårtensson, P., Hanson, L., Björkman, M., Maki, A.: Domain randomization for object detection in manufacturing applications using synthetic data: A comprehensive study. In: International Conference on Robotics and Automation (ICRA). p. 16715–16721. IEEE (2025) 63. Zitkovich, B., Yu, T., Xu, S., Xu, P., Xiao, T., Xia, F., Wu, J., Wohlhart, P., Welker, S., Wahid, A., et al.: Rt-2: Vision-language-action models transfer web knowledge to robotic control. In: Conference on Robot Learning, CoRL. PMLR (2023) Appendix A Implementation Details A.1 Model Architecture Details The model architecture is based on OpenVLA-OFT [17]. In the original Open- VLA [18] framework, a single image and a language instruction are jointly pro- cessed to predict robot actions. Visual features are extracted by a fused vi- sion backbone composed of SigLIP [57] and DINOv2 [38] vision transformers, projected into the language embedding space through an MLP projector, and concatenated with language embeddings before being processed by the Llama- 2 decoder [50]. Building on OpenVLA, OpenVLA-OFT further projects robot proprioceptive state into the language embedding space using a 2-layer MLP with GELU activation, replaces causal attention with bidirectional attention for parallel decoding, substitutes the original language model output head with a 4-layer MLP with ReLU activation to directly predict continuous actions instead of discrete tokens, and predicts chunks of K actions at each step rather than a single-timestep action. In addition, FiLM [40] modules are introduced to mod- ulate visual features in both the SigLIP and DINOv2 vision transformers using the averaged task language embedding. The complete architecture is illustrated in Fig. 4. Architecture Adaptation for RL finetuning. For the RL experiments, we follow SimpleVLA [19] and retain only the parallel decoding and action chunking designs. Instead of replacing the language model head with an MLP for contin- uous action regression, the model keeps the original LLaMA2 output head to generate discrete action tokens and is trained with a cross-entropy loss. This modification enables efficient rollout by allowing the policy to sample actions as discrete tokens using standard language-model decoding procedures. The su- pervised fine-tuning stage uses the same hyperparameter settings as the official OpenVLA implementation. A.2 Training Details SFT and RL fine-tuning are conducted on 8×NVIDIA H100 80GB GPUs. Specif- ically, the SFT stage employs LoRA with rank 32, a batch size of 64, and an initial learning rate of 5× 10 −4 , with learning rate decay starting after 15,000 gradient steps, and the total number of training steps varies from 20,000 to 30,000 depending on the task horizon. The RL fine-tuning stage is initialized from the SFT checkpoint, with a training batch size of 64, a validation batch size of 256, and 8 rollout samples collected per prompt. The actor learning rate is set to 5× 10 −6 with constant warmup. The PPO mini-batch size is 128, the trajectory mini-batch size is 8, and gradient clipping is set to 1. The PPO clip- ping thresholds are ε L = 0.2 and ε H = 0.28, and the rollout temperature is set 1 2F. Author et al. Fig. 1: OpenVLA-OFT+ on the bimanual ALOHA robot.Our Optimized Fine-Tuning (OFT) recipe enhances fine-tuned OpenVLA policies through improved inference efficiency, model quality, and input-output flexibility. The resulting OpenVLA-OFT+ policies execute diverse dexterous manipulation tasks on a real-world bimanual robot at high control frequencies (25 Hz). The “+” suffix indicates the integration of feature-wise linear modulation (FiLM) [37], which strengthens language grounding in tasks where accurate language understanding is critical for success. denoting the augmented recipe asOFT+. OpenVLA-OFT+ successfully executes dexterous bimanual tasks like folding clothes and manipulating target food items based on the user’s prompt (see Figure 1), outperforming both fine-tuned VLAs (π 0 and RDT-1B) and prominent imitation learning policies trained from scratch (Diffusion Policy and ACT) by up to 15% (absolute) in average success rate. With 25-timestep action chunks, OpenVLA-OFT+ achieves 43×faster throughput than base OpenVLA, demonstrating that our new fine-tuning recipe enables real-time robot control with strong task performance and language following ability. I. RELATEDWORK Prior works have leveraged language and vision founda- tion models to enhance robotic capabilities, using them as pretrained visual representations that accelerate robotic policy learning [30, 33, 31, 20, 32], for object localization in robotics tasks [9, 47], and for high-level planning and reasoning [1, 17, 44, 16, 45, 18, 6]. More recently, researchers have explored fine-tuning vision-language models (VLMs) to directly predict low-level robotic control actions, producing “vision-language- action” models (VLAs) [4, 34, 24, 23, 7, 15, 8, 53, 58, 54, 3, 2], which have demonstrated effective language following [59] and generalization to out-of-distribution test conditions and unseen semantic concepts. These works focus primarily on model development, while we focus on developing a recipe for fine-tuning such models, justifying individual design decisions with insights that we gain from our empirical analysis. Despite the importance of fine-tuning for real-world VLA deployment, empirical analysis of effective fine-tuning recipes remains limited. While Kim et al. [23] study various parameter update strategies and from their findings show that LoRA fine- tuning enables effective adaptation to single-arm robots oper- ating at low control frequencies (<10Hz), their analysis does not extend to bimanual robots with high control frequencies (25-50+ Hz), a more complex control scenario. We address this gap by exploring VLA adaptation design decisions for fast inference and reliable task execution on a real-world bimanual manipulator with a 25 Hz controller. Recent works by Belkhale and Sadigh [2] and Pertsch et al. [38] improve VLA efficiency through new action to- kenization schemes, using vector quantization or discrete co- sine transform-based compression to represent action chunks (sequences of actions) with fewer tokens than simple per- dimension binning (as used in RT-2 [4] and OpenVLA [23]). While these approaches achieve 2 to 13×speedups for autoregressive VLAs, we explore design decisions beyond autoregressive modeling, which remains inherently limited by iterative generation. Our parallel decoding approach, when paired with action chunking, achieves significantly greater speedups: 26×to 43×throughput with much lower latency (0.07 ms for single-arm tasks with one input image and 0.321 ms for bimanual tasks with three input images). Another line of research [54, 27, 3] demonstrates effective VLA fine-tuning for high-frequency, bimanual manipulation using generative approaches like diffusion or flow match- Fig. 4: Model architecture of OpenVLA-OFT. to 1.6. We use an action token length of 14 and an action chunk length of 25. The maximum prompt length is 256 and the maximum response length is 128. RL training runs for 50 epochs, using the GRPO advantage estimator with zero KL regularization. To improve training efficiency, rollout samples are filtered by accuracy, retaining only those with success rates between 0.1 and 0.9. B Simulation Datasets B.1 Task Description We evaluate our method on five bimanual manipulation tasks from the RoboTwin 2.0 benchmark [4]. Tab. 4 summarizes the task descriptions and their average step counts. Table 4: Task descriptions and average step counts for the five evaluation tasks. Task NameDescriptionAvg. Steps Stack Bowls Two Stack two bowls on top of each other313 Beat Block Hammer Use the hammer to strike the block.113 Pick Dual Bottles Pick up one bottle with each arm.127 Click BellClick the bell’s top center on the table.85 Place Empty Cup Use an arm to place the empty cup on the coaster.174 B.2 Domain randomization Details Domain randomization includes cluttered scenes, random lighting, table height variation, unseen language instructions, randomized background textures, and head camera pose perturbation. Title Suppressed Due to Excessive Length3 – Scene Clutter: Randomly add task-irrelevant distractors from RoboTwin- OD (731 objects, 147 categories) with collision-aware placement. – Background Textures: Apply random textures from a curated library of 11,000 high-quality surface. – Lighting: Randomize light color, type, intensity, and position within phys- ically plausible bounds. – Tabletop Height: Uniformly vary table height up to ±3cm. – Camera Pose: Apply random displacement up to 0.01m from nominal mounting position. B.3 Fedility Rendering Details Photorealistic Rendering Details. We configure three levels of rendering fidelity to systematically vary visual realism and computational cost across ex- periments: – High fidelity: Ray-tracing renderer with 32 samples per pixel (spp), ray- tracing path depth of 8, and OIDN denoising enabled. – Medium fidelity: Ray-tracing renderer with 4 spp, path depth of 4, and denoiser disabled. – Low fidelity: Standard (non-ray-tracing) renderer with 4 spp and path depth of 4, and denoiser disabled. Physical-Realistic Rendering Details. We construct an intentionally ex- treme simulation environment to evaluate the sensitivity of VLA policies to phys- ical dynamics. Specifically, friction coefficients are reduced to 0.001 (from 0.5), restitution is increased to 0.2 (from 0.0), and gravity is lowered to 0.05 m/s 2 (from 9.81 m/s 2 ). These values create a substantial deviation from real-world physics, producing object behaviors that differ noticeably from those encountered in re- alistic environments. Rather than modeling plausible physical conditions, this setup serves as a controlled stress test that amplifies dynamic discrepancies, al- lowing us to systematically examine how sensitive VLA policies are to changes in environmental dynamics and to better characterize their robustness under severe Sim-to-Real mismatch. C Further Results C.1 Real-World Evaluation Results All real-world experiments are conducted under a canonical base configuration to provide a controlled evaluation setup. In this base setting, the scene uses a white background, no lighting variation, the same object instances as in training, and no table distractors. This configuration serves as the reference environment for evaluating the baseline performance of the policy. To evaluate the robustness of the policy under different environmental con- ditions, we introduce several variations including background changes, lighting 4F. Author et al. Click BellPlace Empty CupBeat Block HammerStack Bowls TwoPick Dual Bottles Fig. 5: Pose design for real-world evaluation across five manipulation tasks. The Place Empty Cup task includes eight pose variations, while the other tasks use four poses variations. Participant Real-World Evaluation Our Site Base Model (Sim Data + Sim2Real Techniques) Training Locally Upload Policy Fig. 6: Overview of the Sim-to-Real benchmark platform. Participants train policies locally using simulation data and Sim2Real techniques, upload the trained models, and the policies are evaluated under standardized real-world robotic setups at our site. variation, object replacement, and the presence of table distractors. Each varia- tion is evaluated independently to analyze how different factors affect real-world performance. For every evaluation condition, the object placement is systemat- ically varied using a predefined 3× 3 grid layout(see Fig. 5), which changes the spatial position of the object relative to the robot. Tabs. 5 to 8 report detailed real-world evaluation results across different real- world environments, analyzing how policy performance changes with variations in visual appearance, physical fidelity, and RL fine-tuning. Among the evaluated factors, distractors introduce the largest performance degradation, while lighting variation has the smallest effect on success rate. D Sim-to-Real Benchmark Platform Inspired by the RoboChallenge [54] real-robot evaluation framework, we design a platform that enables researchers to evaluate policies trained in simulation on real robotic systems. As illustrated in Fig. 6, participants start from a base model and perform local training using simulated environments while applying Sim2Real techniques to improve transferability. The resulting policy is then up- loaded to our centralized evaluation site, where it is executed on real robots under standardized experimental setups. This design allows researchers to de- velop and iterate policies locally without requiring direct access to hardware, while ensuring consistent and reproducible real-world evaluation across different methods. Title Suppressed Due to Excessive Length5 Table 5: Factorized domain randomization results under zero-shot real-world evalua- tion on five tasks. TaskData setting Real setting BaseBGLight Distractor Object Click Bell Clean1/201/603/600/602/60 BG6/208/606/604/606/60 LT6/205/6010/604/607/60 TD2/201/601/603/601/60 CP10/20 12/60 13/609/6017/60 TH14/20 20/60 22/6016/6024/60 TH + CP + BG 7/20 13/60 14/6010/6011/60 TH + CP + LT 7/20 11/60 14/609/6011/60 TH + CP + TD 6/208/609/6010/6010/60 All Factors15/20 30/60 39/60 14/60 31/60 Place Empty Cup Clean3/40 9/120 7/1202/1207/120 BG8/40 16/120 16/1203/12010/120 LT6/40 9/120 17/1203/12010/120 TD4/40 8/120 12/1207/1205/120 CP13/40 12/120 35/120 12/120 22/120 TH16/40 24/120 43/120 14/120 32/120 TH + CP + BG 23/40 42/120 51/120 23/120 35/120 TH + CP + LT 19/40 29/120 50/120 12/120 35/120 TH + CP + TD 21/40 28/120 32/120 26/120 30/120 All Factors22/40 42/120 53/120 38/120 58/120 Beat Block Hammer Clean0/200/600/600/600/60 BG1/203/602/600/601/60 LT2/203/606/600/600/60 TD0/200/600/600/600/60 CP2/204/604/600/602/60 TH3/204/606/600/604/60 TH + CP + BG 4/206/607/600/605/60 TH + CP + LT 3/209/609/600/603/60 TH + CP + TD 3/204/605/600/605/60 All Factors3/20 9/60 10/60 0/608/60 Stack Bowls Two Clean9/209/6019/6012/6019/60 BG9/20 24/60 29/6016/6027/60 LT9/20 18/60 23/6013/6019/60 TD8/20 12/60 17/6018/6016/60 CP11/20 23/60 31/6017/6028/60 TH12/20 30/60 33/6019/6035/60 TH + CP + BG 13/20 38/60 40/6028/6037/60 TH + CP + LT 13/20 32/60 35/6025/6037/60 TH + CP + TD 12/20 33/60 32/6026/6034/60 All Factors14/20 40/60 42/60 29/60 39/60 Pick Dual Bottles Clean1/201/601/600/601/60 BG4/20 10/607/605/606/60 LT3/203/606/603/604/60 TD2/202/602/604/602/60 CP5/209/6010/608/6010/60 TH5/209/6011/606/609/60 TH + CP + BG 7/20 13/60 14/60 10/6011/60 TH + CP + LT 7/20 11/60 14/609/6011/60 TH + CP + TD 6/208/609/6010/6010/60 All Factors7/20 15/60 17/60 10/60 13/60 6F. Author et al. Table 6: Frame-wise domain randomization results under zero-shot real-world evalu- ation across five tasks. TaskData setting Real setting Base BG Light Distractor Object Click Bell Frame-wise CP 10/20 16/60 22/6012/6023/60 Frame-wise BG 4/20 11/60 8/606/6010/60 Frame-wise LT 4/20 6/60 12/605/6011/60 Place Empty Cup Frame-wise CP 18/40 24/120 39/120 19/120 33/120 Frame-wise BG 10/40 29/120 30/120 23/120 28/120 Frame-wise LT 6/40 9/120 19/120 3/120 10/120 Beat Block Hammer Frame-wise CP 3/20 5/60 6/600/605/60 Frame-wise BG 3/20 6/60 5/600/606/60 Frame-wise LT 2/20 4/60 5/600/602/60 Stack Bowls Two Frame-wise CP 15/20 25/60 34/6022/6032/60 Frame-wise BG 10/20 30/60 31/6017/6029/60 Frame-wise LT 9/20 18/60 26/6013/6019/60 Pick Dual Bottles Frame-wise CP 6/20 13/60 14/607/6011/60 Frame-wise BG 4/20 12/60 7/605/609/60 Frame-wise LT 3/20 2/60 8/602/604/60 Table 7: Impact of photorealistic and physical-realistic fidelity on zero-shot real-world policy generalization. Default denotes high-fidelity rendering with ray-tracing enabled and standard physics settings. All experiments are trained using the All Factors domain randomization configuration. TaskData setting Real setting BaseBGLight Distractor Object Click Bell Default15/20 30/60 39/60 14/60 31/60 Medium Visual 14/20 30/60 32/60 14/6030/60 Low Visual3/207/609/605/6010/60 Extreme Physics 12/20 24/60 32/6011/6026/60 Place Empty Cup Default22/40 42/120 53/120 38/120 58/120 Medium Visual 12/40 23/120 30/120 20/120 35/120 Low Visual8/40 15/120 19/120 13/120 21/120 Extreme Physics 20/40 37/120 46/120 34/120 51/120 Beat Block Hammer Default3/20 9/60 10/600/608/60 Medium Visual 3/20 7/605/600/605/60 Low Visual0/200/600/600/600/60 Extreme Physics 1/202/602/600/602/60 Stack Bowls Two Default14/20 40/60 42/60 29/60 39/60 Medium Visual 12/20 32/60 34/6028/6037/60 Low Visual10/20 24/60 27/6021/6027/60 Extreme Physics 14/20 37/60 41/60 29/6033/60 Pick Dual Bottles Default7/20 15/60 17/60 10/60 13/60 Medium Visual 5/20 12/60 13/608/6012/60 Low Visual2/204/604/602/604/60 Extreme Physics 3/207/608/605/608/60 Title Suppressed Due to Excessive Length7 Table 8: RL fine-tuning results under zero-shot real-world evaluation across five tasks. TaskData setting Real setting BaseBGLight Distractor Object Click Bell SFT0/200/600/600/600/60 SFT + RL11/20 26/60 28/6012/6019/60 SFT + RL + DR 13/20 32/60 35/60 23/60 29/60 Place Empty Cup SFT6/40 7/120 12/1203/1207/120 SFT + RL19/40 32/120 52/120 24/120 54/120 SFT + RL + DR 26/40 59/120 75/120 43/120 64/120 Beat Block Hammer SFT0/200/600/600/600/60 SFT + RL4/208/608/600/605/60 SFT + RL + DR 5/20 12/60 14/60 1/60 10/60 Stack Bowls Two SFT6/20 13/60 14/609/6013/60 SFT + RL14/20 35/60 41/6028/6036/60 SFT + RL + DR 16/20 38/60 45/60 31/60 38/60 Pick Dual Bottles SFT0/200/600/600/600/60 SFT + RL7/20 16/60 20/6012/6014/60 SFT + RL + DR 8/20 20/60 21/60 16/60 15/60