Pith. sign in

REVIEW 4 major objections 6 minor 2 cited by

The paper claims that a two-stage recipe—supervised fine-tuning on mixed real and simulated data, then simulation reinforcement learning with a real-data anchor—improves real-robot success, generalization, and data efficiency for vision-lan

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 23:44 UTC pith:2LJSLGYN

load-bearing objection A practical two-stage sim-RL co-training recipe for VLAs with clean ablations; the 'consistent substantial gains' claim needs more than two runs per cell. the 4 major comments →

arxiv 2602.12628 v4 pith:2LJSLGYN submitted 2026-02-13 cs.RO

Beyond Imitation: Reinforcement Learning-Based Sim-Real Co-Training for VLA Models

Liangzhi Shi , Shuaihang Chen , Feng Gao , Yinuo Chen , Kang Chen , Tonghe Zhang , Hongzhi Zang , Jiakai Zhou
show 3 more authors
Weinan Zhang Chao Yu Yu Wang
This is my paper
classification cs.RO
keywords vision-language-action (VLA) modelsreinforcement learning fine-tuningsim-to-real transfersim-real co-trainingrobot manipulationcatastrophic forgettingdata efficiencygeneralization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that a vision-language-action model can become more capable on a real robot through a two-stage training scheme: first imitate a mixture of real and simulated demonstrations, then run reinforcement learning in a low-fidelity simulator while continuing to supervise on a small set of real demonstrations. The central claim is that this RL-based sim-real co-training consistently beats both real-only fine-tuning and supervised co-training, with average real-world success gains of 24 percentage points on OpenVLA and 20 on π0.5 across four tabletop tasks. It also claims stronger robustness to unseen objects and robot states, and a large cut in the amount of real demonstration data required. If true, the recipe offers a practical, model-agnostic way to exploit cheap interactive simulation for real-robot deployment without photorealistic simulators or real-world RL.

Core claim

The paper's discovery is that the missing ingredient in sim-real co-training is not more simulation data but closed-loop interaction paired with a real-world anchor. RL-Co first initializes the policy by supervised fine-tuning on a mixture of real and simulated demonstrations, then fine-tunes it with reinforcement learning in a digital-twin simulator while minimizing a total loss that adds a supervised term on real data. That real-data term is what stops catastrophic forgetting; removing it drops the real-world success on one task from 81.3% to 40.3%. In evaluations on four real-world tabletop tasks with OpenVLA (a next-token-prediction VLA) and π0.5 (a flow-matching VLA, a generative action

What carries the argument

The load-bearing mechanism is the two-stage loss combination. Stage I applies supervised co-training with a mixture ratio α, L_SFT = αL_SFT(D_sim) + (1−α)L_SFT(D_real), to give the policy a competent start in both the real and simulated versions of the task. Stage II optimizes L_total = L_RL + βL_SFT(D_real): the reinforcement-learning loss drives exploration and skill improvement in the low-fidelity simulator, while the auxiliary supervised loss on real demonstrations anchors the policy to human behavior and prevents catastrophic forgetting. The recipe is architecture-agnostic: the same total objective is applied to a next-token-prediction VLA (OpenVLA) and a flow-matching VLA (π0.5), using

Load-bearing premise

The load-bearing premise is that skills learned in the simplified simulator—built with only the essential object meshes and geometry, without realistic materials or lighting—transfer to the real robot well enough to improve real-world success beyond what imitation co-training achieves.

What would settle it

Train RL-Co with the simulator's physics deliberately mismatched from the real robot (for example, object masses doubled, friction reversed, gripper latency added) and compare real-world success against SFT-based co-training. If the real-world gains disappear or turn negative as the sim-to-real gap widens, the method's success depends on a transfer assumption the paper does not quantify.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Real-world success rates rise on every tested task and model, averaging +24 percentage points (OpenVLA) and +20 (π0.5) over SFT-based sim-real co-training.
  • Generalization under distribution shift improves: on the Pick and Place task with π0.5, RL-Co degrades by 25 points on unseen objects where real-only training degrades by 47 and SFT co-training by 38.
  • Real-world data requirements drop sharply: on Open Drawer with π0.5, RL-Co trained with 20 demonstrations performs comparably to or better than SFT co-training trained with 200.
  • The real-data anchor in Stage II is necessary: removing it drops real-world success on Pick and Place from 81.3% to 40.3%, so the RL gain depends on anchoring, not just on simulation RL.
  • Because the same objective works with two different VLA action decoders, the recipe is portable to other VLA architectures.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper: the same real-data anchor should make the recipe work with 'digital cousins' rather than exact digital twins, since the paper's simulators deliberately omit materials and lighting; a natural test is co-training on tasks that share only structure, not object geometry.
  • Beyond the paper: because the anchor keeps the policy close to demonstrated behavior, Stage II could be interleaved with occasional real-world RL rollouts without the usual catastrophic forgetting—the paper lists real-world RL as future work and does not test this.
  • Beyond the paper: the regularization weight β varies by task in the paper's hyperparameter table, suggesting that annealing β from high to low as RL competence grows could reduce tuning burden and is a testable extension.
  • Beyond the paper: the data-efficiency result hints that the method could cut real demonstration collection by an order of magnitude on other manipulation families, but the paper only evaluates four single-arm tabletop tasks, so that scope is unproven.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes RL-Co, a two-stage sim-real co-training recipe for vision-language-action (VLA) models. Stage I warm-starts the policy with SFT on a mixture of real and simulated demonstrations; Stage II runs RL in a ManiSkill digital twin while adding an auxiliary SFT loss on real-world demonstrations to prevent catastrophic forgetting. The method is evaluated on four tabletop tasks (Pick and Place, Push Cube, Open Drawer, Close Drawer) with OpenVLA and pi0.5, reporting higher real-world success than real-only SFT and SFT-based co-training, plus generalization and data-efficiency gains. The paper releases code, checkpoints, and a project page.

Significance. If the headline results hold, RL-Co would be a practical, architecture-agnostic recipe for using closed-loop simulation in VLA post-training, and the ablation design (sim-SFT initialization, real-SFT in both stages, alpha/beta sensitivity) is informative. The contribution is incremental over existing RL fine-tuning and sim-real co-training pipelines, but the real-robot evaluation across two VLA families is a useful data point. The paper also ships reproducible assets, which is a clear strength. However, the central claim of consistent, substantial real-world gains is currently supported only by n=2 runs per cell, with several pairwise differences inside sampling noise, so the empirical evidence is not yet commensurate with the strength of the claims.

major comments (4)
  1. [§V-A.1, Table I] The central claim "RL-Co consistently yields substantially higher real-world success rates across all task and model combinations" is not statistically supported by the reported evaluation. The text states "Each setting is evaluated twice," so Table I's mean±std is based on n=2 training runs per cell. Several advertised gains are within sampling noise: pi0.5 Pick and Place 81.3±9.4 vs 68.8±9.4 (Δ=12.5, se≈9.4); OpenVLA Push Cube 68.3±11.7 vs 51.7±5.0 (Δ=16.6, se≈9.0); OpenVLA Close Drawer 95.0±5.0 vs 85.0±5.0 (Δ=10, se=5). No significance test, effect size, or confidence interval is reported, and the number of evaluation episodes per run is not given. The "consistent" and "substantial" wording therefore overstates the evidence. Please provide more seeds, report episode counts, and add a paired or bootstrap analysis.
  2. [§V-B, Table IV] The task-specific hyperparameters alpha, beta, and learning rates in Table IV appear to have been tuned individually. The paper does not describe a validation protocol or state whether the same alpha was used for the SFT co-training baseline. If alpha/beta were selected with knowledge of the real-robot evaluation results, the comparison risks capitalizing on chance. Please report how alpha and beta were chosen (e.g., a held-out simulation validation set), and show the sensitivity of the headline comparisons over a range of alpha and beta. This is load-bearing for the claim that RL-Co robustly beats SFT co-training rather than merely outperforming it at favorable hyperparameter settings.
  3. [§V-B, Table II] The generalization claim rests on a single task (Pick and Place), a single model (pi0.5), and point estimates without error bars or significance testing. The in-distribution row is itself from n=2, and the unseen-object/unseen-state numbers appear to come from a single evaluation. The abstract and Section V-B state that RL-Co yields "significantly stronger generalization" and "markedly smaller performance degradation," but the evidence is not statistically grounded. Add repeated evaluations with confidence intervals or soften the claim to a preliminary observation.
  4. [§V-D, Fig. 7] The data-efficiency claim is based on one task (Open Drawer) and does not state the number of training seeds or evaluation episodes. The statement that RL-Co with 20 demonstrations outperforms baselines with 200 demonstrations needs error bars or a statistical comparison; with n=2 the shaded standard deviations are unreliable. Since the abstract highlights "substantially improved real-world data efficiency," this claim needs stronger support than a single-task curve.
minor comments (6)
  1. [Fig. 6] The text reports 81.38% for pi0.5 Pick and Place, while Table I reports 81.3±9.4. Please make precision consistent.
  2. [Table II] The value "31.95↓" is oddly precise and inconsistent with the other entries; if this is a relative drop it should be 44.4%, not 31.95. Please clarify the metric.
  3. [Author affiliations] Typo: "Peking Unviersity" should be "Peking University."
  4. [§V-A.2 / Table IV] The text says 20–50 real demonstrations per task, and Table IV lists 50, 50, 20, 30. Please state the exact counts in the text for each task.
  5. [Fig. 4] The beta values tested are not enumerated in the caption or text. Please list them explicitly so the sensitivity analysis is reproducible.
  6. [Fig. 7] The shaded regions are described as standard deviation but the number of seeds is not stated. Please specify the seed count and evaluation protocol in the figure caption.

Circularity Check

0 steps flagged

No significant circularity: the real-robot success-rate claims are measured outcomes, not derived from fitted parameters or author-uniqueness theorems.

full rationale

The paper's derivation chain is methodological rather than predictive. Stage I (Eq. 7) and Stage II (Eq. 8) define an optimization objective that mixes SFT co-training with an RL term and a real-data regularizer; no quantity in these equations is defined in terms of the reported real-world success rate, and no fitted parameter is later renamed as a prediction. The central claims—consistent real-world gains, improved generalization, and better data efficiency—are supported by direct real-robot measurements (Table I, Table II, Fig. 5–7). Per-task α and β values are tuned hyperparameters, but tuning hyperparameters on training/validation data is model selection, not circularity. The paper cites several prior works by overlapping authors (Liu et al. [43], ReinFlow [81], RLinf [76], πRL [14]) as RL machinery and motivation, but these citations are not load-bearing proof of the central empirical result; the paper's own ablations and evaluations stand independently of them. The acknowledged limitations (single embodiment, tabletop tasks, no real-world RL) are scope constraints, not circular steps. The statistical weakness noted by a skeptical reader—n=2 evaluation runs and lack of significance testing—is a rigor/validity concern, not an equivalence between input and output. I therefore find no self-definitional, fitted-input-as-prediction, or self-citation-driven circularity.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The central claim rests on the transferability of RL-learned behavior from a deliberately non-photorealistic simulator, on the effectiveness of the real-SFT anchor, and on the reliability of very small real-robot evaluation sets. The only tuned quantities are the co-training ratio α, the regularization weight β, and hand-shaped reward coefficients.

free parameters (3)
  • α (Stage I co-training ratio) = 0.5 (OpenVLA); 0.5–0.98 (π0.5, per task)
    Controls fraction of simulated demonstrations in SFT co-training (Eq. 7); tuned per model-task (Table IV), with sensitivity shown in Fig. 4.
  • β (Stage II real-SFT regularization weight) = 0.01–1.0 across tasks/models
    Weights the real-world SFT anchor loss against the RL loss (Eq. 8); tuned per model-task (Table IV), with sensitivity in Fig. 4.
  • Reward shaping coefficients (success bonuses, distance gains, penalties) = e.g., 0.2/1.0 grasps, 3.0/5.0 success bonuses, −0.4 penalty, tanh(5x)/tanh(10x) shaping
    Hand-designed per-task reward functions in Appendix C.1; these determine what RL optimizes and differ between OpenVLA and π0.5.
axioms (4)
  • domain assumption The low-fidelity ManiSkill digital twin is a sufficient proxy for the real setup so that RL improvements transfer to the real robot when anchored by real SFT.
    Section V-A.1 explicitly omits materials/lighting; entire RL-Co Stage II presupposes this transferability.
  • domain assumption Auxiliary SFT loss on real demonstrations preserves real-world capability during simulation RL (catastrophic-forgetting prevention).
    This is the mechanism of Eq. (8), supported only by the ablation in Fig. 6.
  • standard math The adopted RL algorithms (PPO-style for OpenVLA, ReinFlow for π0.5) converge in this setting and their hyperparameter choices are sound.
    Borrowed from prior works [43, 76, 81], not re-derived here.
  • domain assumption Two evaluation runs per setting yield success-rate estimates precise enough to rank methods.
    Section V-A.1; with n=2, reported std is effectively uninformative and many confidence intervals overlap.

pith-pipeline@v1.3.0-alltime-deepseek · 19938 in / 11311 out tokens · 89570 ms · 2026-08-02T23:44:12.339756+00:00 · methodology

0 comments
read the original abstract

Simulation offers a scalable and low-cost way to enrich vision-language-action (VLA) training, reducing reliance on expensive real-robot demonstrations. However, most sim-real co-training methods rely on supervised fine-tuning (SFT), which treats simulation as a static source of demonstrations and does not exploit large-scale closed-loop interaction. Consequently, real-world gains and generalization are often limited. In this paper, we propose an RL-based sim-real Co-training (RL-Co) framework that leverages interactive simulation while preserving real-world capabilities. Our method follows a generic two-stage design: we first warm-start the policy with SFT on a mixture of real and simulated demonstrations, then fine-tune it with reinforcement learning in simulation while adding an auxiliary supervised loss on real-world data to anchor the policy and mitigate catastrophic forgetting. We evaluate our framework on four real-world tabletop manipulation tasks using two representative VLA architectures, OpenVLA and $\pi_{0.5}$, and observe consistent improvements over real-only fine-tuning and SFT-based co-training, including +24% real-world success on OpenVLA and +20% on $\pi_{0.5}$. Beyond higher success rates, RL co-training yields stronger generalization to unseen task variations and substantially improved real-world data efficiency, providing a practical and scalable pathway for leveraging simulation to enhance real-robot deployment.

Figures

Figures reproduced from arXiv: 2602.12628 by Chao Yu, Feng Gao, Hongzhi Zang, Jiakai Zhou, Kang Chen, Liangzhi Shi, Shuaihang Chen, Tonghe Zhang, Weinan Zhang, Yinuo Chen, Yu Wang.

Figure 1
Figure 1. Figure 1: Overview of training paradigms combining real￾world and simulated data. VLA models are commonly trained via supervised fine-tuning (SFT) on real-world demon￾strations, or via reinforcement learning (RL) in simulation followed by sim-to-real transfer. Other approaches adopt SFT￾based sim–real co-training by mixing real and simulated demonstrations. In contrast, we propose an RL-based sim– real co-training (… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed two-stage sim-real co-training framework. We establish a digital-twin setup where Tsim serves as a digital cousin to Treal despite visual discrepancies. In Stage I, we initialize the VLA policy by supervising it on a mixture of real and simulated data (ratio α). This rapidly injects real-world knowledge and prepares the policy for simulation interaction. In Stage II, we perform RL … view at source ↗
Figure 3
Figure 3. Figure 3: Visualization of our tabletop manipulation tasks. The top row shows images captured by a third-person camera in the real-world setup, while the bottom row presents the cor￾responding simulated views. Both real and simulated images are sampled from the task execution. We construct the simulation environments using Man￾iSkill [60], matching the real-world setup in terms of camera viewpoints and scene layout.… view at source ↗
Figure 4
Figure 4. Figure 4: Analysis of the co-training ratio (α) and regulariza￾tion weight (β). We vary the co-training ratio α and evaluate the resulting performance on the Pick and Place and Open Drawer tasks. In addition, we fix α = 0.5 for Pick and Place and α = 0.95 for Open Drawer, reporting RL co-training results under different regularization weights β. Performance is measured by success rate, with shaded regions indicating… view at source ↗
Figure 7
Figure 7. Figure 7: Effect of the number of real-world demonstrations. We vary the amount of real-world demonstrations for the Open Drawer task and evaluate all training paradigms using the π0.5 model. Performance is reported in terms of success rate, with shaded regions indicating the standard deviation. D. Data Efficiency As shown in Section V-B, RL-Co outperforms both real￾only training and SFT-based co-training under the … view at source ↗
Figure 6
Figure 6. Figure 6: Ablation study on real-world supervision. We ablate real-world supervised training in Stage I and Stage II sepa￾rately and report the resulting real-world success rates. reports the real-world success rates of the Pick and Place task using the π0.5 model under all ablation settings. When the real-world SFT regularization is removed from Stage II, the success rate drops significantly from 81.38% to 40.25%, … view at source ↗
Figure 8
Figure 8. Figure 8: illustrates our real-world evaluation setup. The system consists of a table-top workspace, a Franka Emika Panda robot mounted on the table, and a fixed RGB camera. We use the RGB channels captured by the camera as the visual input to the VLA model. Desk Franka Emika Panda robot RGB camera [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Visualization of Four Tabletop Manipulation Tasks. For each task, we present one successful trajectory and uniformly sample seven frames along the execution. Each row corresponds to a single trajectory shown from start to completion. categories: regular-shaped and irregular-shaped. Regular￾shaped objects consist of toy fruits and vegetables, while irregular-shaped objects include bowls and gloves. No￾tably… view at source ↗
Figure 10
Figure 10. Figure 10: Manipulated Objects in Simulation and the Real World. The left panel shows the objects used in simulation, while the right panel presents the real-world objects. All simulated objects are used during training. The real-world objects are divided into training objects and unseen objects for generalization evaluation. • Push Cube: For each evaluation episode, three cubes are randomly selected from all availa… view at source ↗
Figure 11
Figure 11. Figure 11: Initial Regions for Manipulative Objects. For the Pick and Place task, the bowl is placed within the orange region, while the objects are initialized in the blue region. For the Push Cube task, each cube is initialized within its corresponding orange region. For the Open / Close Drawer tasks, the front edge of the drawer is initialized within the orange region. with the drawer orientation initially aligne… view at source ↗
Figure 12
Figure 12. Figure 12: Simulation Training Results. We report the simulation success rates across all settings during RL training. Parameter Names Setting Pick and Place Push Cube Open Drawer Close Drawer General Number of Real Demos 50 50 20 30 OpenVLA Co-training ratio α 0.5 0.5 0.5 0.5 SFT learning rate 5 × 10−4 5 × 10−4 5 × 10−4 5 × 10−4 Regularization weight β 0.1 0.01 0.01 0.01 Actor learning rate 10−4 10−4 10−4 10−4 Crit… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. WCM: A World Critic Model for Vision-Language-Action Reinforcement Learning

    cs.RO 2026-07 conditional novelty 6.0

    A critic that jointly predicts future latent states and values improves RL fine-tuning and out-of-distribution generalization for vision-language-action robot policies.

  2. TacCoRL: Integrating Tactile Feedback into VLA via Simulation

    cs.RO 2026-06 unverdicted novelty 5.0

    TacCoRL integrates tactile feedback into VLA policies via real-aligned simulation co-training and RL, raising average success from 50% to 72.5% on four bimanual contact-rich tasks with direct real-robot transfer.

Reference graph

Works this paper leans on

91 extracted references · 44 linked inside Pith · cited by 2 Pith papers

  1. [1]

    Learning dexterous in-hand manipula- tion.The International Journal of Robotics Research, 39 (1):3–20, 2020

    OpenAI: Marcin Andrychowicz, Bowen Baker, Maciek Chociej, Rafal Jozefowicz, Bob McGrew, Jakub Pa- chocki, Arthur Petron, Matthias Plappert, Glenn Powell, Alex Ray, et al. Learning dexterous in-hand manipula- tion.The International Journal of Robotics Research, 39 (1):3–20, 2020

    2020

  2. [2]

    From imitation to refinement- residual rl for precise assembly

    Lars Ankile, Anthony Simeonov, Idan Shenfeld, Marcel Torne, and Pulkit Agrawal. From imitation to refinement- residual rl for precise assembly. In2025 IEEE Interna- tional Conference on Robotics and Automation (ICRA), pages 01–08. IEEE, 2025

    2025

  3. [3]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wen- bin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report.arXiv preprint arXiv:2502.13923, 2025

    Pith/arXiv arXiv 2025

  4. [4]

    Efficient online reinforcement learning with offline data

    Philip J Ball, Laura Smith, Ilya Kostrikov, and Sergey Levine. Efficient online reinforcement learning with offline data. InInternational Conference on Machine Learning, pages 1577–1594. PMLR, 2023

    2023

  5. [5]

    Paligemma: A versatile 3b vlm for transfer.arXiv preprint arXiv:2407.07726, 2024

    Lucas Beyer, Andreas Steiner, Andr ´e Susano Pinto, Alexander Kolesnikov, Xiao Wang, Daniel Salz, Maxim Neumann, Ibrahim Alabdulmohsin, Michael Tschannen, Emanuele Bugliarello, et al. Paligemma: A versatile 3b vlm for transfer.arXiv preprint arXiv:2407.07726, 2024

    Pith/arXiv arXiv 2024

  6. [6]

    The r2r framework: Publishing and discovering mappings on the web.COLD, 665:97–108, 2010

    Christian Bizer and Andreas Schultz. The r2r framework: Publishing and discovering mappings on the web.COLD, 665:97–108, 2010

    2010

  7. [7]

    Gr00t n1: An open foundation model for generalist humanoid robots.arXiv preprint arXiv:2503.14734, 2025

    Johan Bjorck, Fernando Casta ˜neda, Nikita Cherniadev, Xingye Da, Runyu Ding, Linxi Fan, Yu Fang, Dieter Fox, Fengyuan Hu, Spencer Huang, et al. Gr00t n1: An open foundation model for generalist humanoid robots.arXiv preprint arXiv:2503.14734, 2025

    Pith/arXiv arXiv 2025

  8. [8]

    Kevin Black, Noah Brown, Danny Driess, Adnan Es- mail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al.pi 0: A vision-language-action flow model for general robot control.arXiv preprint arXiv:2410.24164, 2024

    Pith/arXiv arXiv 2024

  9. [9]

    Rt-1: Robotics transformer for real-world control at scale.arXiv preprint arXiv:2212.06817, 2022

    Anthony Brohan, Noah Brown, Justice Carbajal, Yev- gen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale.arXiv preprint arXiv:2212.06817, 2022

    Pith/arXiv arXiv 2022

  10. [10]

    The ycb object and model set: Towards common benchmarks for manipulation research

    Berk Calli, Arjun Singh, Aaron Walsman, Siddhartha Srinivasa, Pieter Abbeel, and Aaron M Dollar. The ycb object and model set: Towards common benchmarks for manipulation research. In2015 international conference on advanced robotics (ICAR), pages 510–517. IEEE, 2015

    2015

  11. [11]

    Shapenet: An information-rich 3d model repository.arXiv preprint arXiv:1512.03012, 2015

    Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository.arXiv preprint arXiv:1512.03012, 2015

    Pith/arXiv arXiv 2015

  12. [12]

    Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual con- cepts, 2021

    Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual con- cepts, 2021. URL https://arxiv.org/abs/2102.08981

    Pith/arXiv arXiv 2021

  13. [13]

    Closing the sim-to-real loop: Adapting simulation randomization with real world experience

    Yevgen Chebotar, Ankur Handa, Viktor Makoviychuk, Miles Macklin, Jan Issac, Nathan Ratliff, and Dieter Fox. Closing the sim-to-real loop: Adapting simulation randomization with real world experience. In2019 International Conference on Robotics and Automation (ICRA), pages 8973–8979. IEEE, 2019

    2019

  14. [14]

    URL https: //arxiv.org/abs/2510.25889

    Kang Chen, Zhihao Liu, Tonghe Zhang, Zhen Guo, Si Xu, Hao Lin, Hongzhi Zang, Xiang Li, Quanlu Zhang, Zhaofei Yu, Guoliang Fan, Tiejun Huang, Yu Wang, and Chao Yu.π RL: Online rl fine-tuning for flow- based vision-language-action models, 2026. URL https: //arxiv.org/abs/2510.25889

    arXiv 2026

  15. [15]

    Sharegpt4v: Improving large multi-modal models with better cap- tions, 2023

    Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better cap- tions, 2023. URL https://arxiv.org/abs/2311.12793

    Pith/arXiv arXiv 2023

  16. [16]

    Robotwin 2.0: A scalable data generator and benchmark with strong domain random- ization for robust bimanual robotic manipulation.arXiv preprint arXiv:2506.18088, 2025

    Tianxing Chen, Zanxin Chen, Baijun Chen, Zijian Cai, Yibin Liu, Zixuan Li, Qiwei Liang, Xianliang Lin, Yi- heng Ge, Zhenyu Gu, et al. Robotwin 2.0: A scalable data generator and benchmark with strong domain random- ization for robust bimanual robotic manipulation.arXiv preprint arXiv:2506.18088, 2025

    Pith/arXiv arXiv 2025

  17. [17]

    Generalizable domain adaptation for sim-and-real policy co-training

    Shuo Cheng, Liqian Ma, Zhenyang Chen, Ajay Man- dlekar, Caelan Garrett, and Danfei Xu. Generalizable domain adaptation for sim-and-real policy co-training. arXiv preprint arXiv:2509.18631, 2025

    arXiv 2025

  18. [18]

    Automated creation of digital cousins for robust policy learning.arXiv preprint arXiv:2410.07408, 2024

    Tianyuan Dai, Josiah Wong, Yunfan Jiang, Chen Wang, Cem Gokmen, Ruohan Zhang, Jiajun Wu, and Li Fei-Fei. Automated creation of digital cousins for robust policy learning.arXiv preprint arXiv:2410.07408, 2024

    Pith/arXiv arXiv 2024

  19. [19]

    Objaverse-xl: A universe of 10m+ 3d objects.Advances in Neural Information Processing Systems, 36:35799– 35813, 2023

    Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram V oleti, Samir Yitzhak Gadre, et al. Objaverse-xl: A universe of 10m+ 3d objects.Advances in Neural Information Processing Systems, 36:35799– 35813, 2023

    2023

  20. [20]

    Objaverse: A universe of annotated 3d objects

    Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objaverse: A universe of annotated 3d objects. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13142–13153, 2023

    2023

  21. [21]

    Challenges of real-world reinforcement learning

    Gabriel Dulac-Arnold, Daniel Mankowitz, and Todd Hes- ter. Challenges of real-world reinforcement learning. arXiv preprint arXiv:1904.12901, 2019

    Pith/arXiv arXiv 1904

  22. [22]

    Bridge data: Boosting generalization of robotic skills with cross- domain datasets, 2021

    Frederik Ebert, Yanlai Yang, Karl Schmeckpeper, Bernadette Bucher, Georgios Georgakis, Kostas Dani- ilidis, Chelsea Finn, and Sergey Levine. Bridge data: Boosting generalization of robotic skills with cross- domain datasets, 2021. URL https://arxiv.org/abs/2109. 13396

    2021

  23. [23]

    Sim-and-human co-training for data- efficient and generalizable robotic manipulation, 2026

    Kaipeng Fang, Weiqing Liang, Yuyang Li, Ji Zhang, Pengpeng Zeng, Lianli Gao, Jingkuan Song, and Heng Tao Shen. Sim-and-human co-training for data- efficient and generalizable robotic manipulation, 2026. URL https://arxiv.org/abs/2601.19406

    arXiv 2026

  24. [24]

    Maniskill2: A unified benchmark for generalizable manipulation skills.arXiv preprint arXiv:2302.04659, 2023

    Jiayuan Gu, Fanbo Xiang, Xuanlin Li, Zhan Ling, Xiqiang Liu, Tongzhou Mu, Yihe Tang, Stone Tao, Xinyue Wei, Yunchao Yao, et al. Maniskill2: A unified benchmark for generalizable manipulation skills.arXiv preprint arXiv:2302.04659, 2023

    Pith/arXiv arXiv 2023

  25. [25]

    Airbert: In-domain pretraining for vision-and-language navigation

    Pierre-Louis Guhur, Makarand Tapaswi, Shizhe Chen, Ivan Laptev, and Cordelia Schmid. Airbert: In-domain pretraining for vision-and-language navigation. InPro- ceedings of the IEEE/CVF international conference on computer vision, pages 1634–1643, 2021

    2021

  26. [26]

    Towards learning a generic agent for vision-and-language navigation via pre-training

    Weituo Hao, Chunyuan Li, Xiujun Li, Lawrence Carin, and Jianfeng Gao. Towards learning a generic agent for vision-and-language navigation via pre-training. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13137–13146, 2020

    2020

  27. [27]

    Vln bert: A recurrent vision- and-language bert for navigation

    Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez- Opazo, and Stephen Gould. Vln bert: A recurrent vision- and-language bert for navigation. InProceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 1643–1653, 2021

    2021

  28. [28]

    Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022

    2022

  29. [29]

    Physical Intelligence, Ali Amin, Raichelle Aniceto, Ash- win Balakrishna, Kevin Black, Ken Conley, Grace Con- nors, James Darpinian, Karan Dhabalia, Jared DiCarlo, et al.π ∗ 0.6: a vla that learns from experience.arXiv preprint arXiv:2511.14759, 2025

    Pith/arXiv arXiv 2025

  30. [30]

    pi {0.5}: a vision-language-action model with open- world generalization.arXiv preprint arXiv:2504.16054, 2025

    Physical Intelligence, Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, Adnan Es- mail, Michael Equi, Chelsea Finn, Niccolo Fusai, et al. pi {0.5}: a vision-language-action model with open- world generalization.arXiv preprint arXiv:2504.16054, 2025

    Pith/arXiv arXiv 2025

  31. [31]

    Vima: General robot manipulation with multimodal prompts.arXiv preprint arXiv:2210.03094, 2(3):6, 2022

    Yunfan Jiang, Agrim Gupta, Zichen Zhang, Guanzhi Wang, Yongqiang Dou, Yanjun Chen, Li Fei-Fei, Anima Anandkumar, Yuke Zhu, and Linxi Fan. Vima: General robot manipulation with multimodal prompts.arXiv preprint arXiv:2210.03094, 2(3):6, 2022

    Pith/arXiv arXiv 2022

  32. [32]

    Qt-opt: Scalable deep reinforcement learning for vision-based robotic manipulation, 2018

    Dmitry Kalashnikov, Alex Irpan, Peter Pastor, Julian Ibarz, Alexander Herzog, Eric Jang, Deirdre Quillen, Ethan Holly, Mrinal Kalakrishnan, Vincent Vanhoucke, and Sergey Levine. Qt-opt: Scalable deep reinforcement learning for vision-based robotic manipulation, 2018. URL https://arxiv.org/abs/1806.10293

    Pith/arXiv arXiv 2018

  33. [33]

    3d gaussian splatting for real-time radiance field rendering.ACM Trans

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering.ACM Trans. Graph., 42(4):139– 1, 2023

    2023

  34. [34]

    Droid: A large-scale in-the-wild robot manipulation dataset.arXiv preprint arXiv:2403.12945, 2024

    Alexander Khazatsky, Karl Pertsch, Suraj Nair, Ash- win Balakrishna, Sudeep Dasari, Siddharth Karam- cheti, Soroush Nasiriany, Mohan Kumar Srirama, Lawrence Yunliang Chen, Kirsty Ellis, et al. Droid: A large-scale in-the-wild robot manipulation dataset.arXiv preprint arXiv:2403.12945, 2024

    Pith/arXiv arXiv 2024

  35. [35]

    Openvla: An open-source vision-language-action model.arXiv preprint arXiv:2406.09246, 2024

    Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model.arXiv preprint arXiv:2406.09246, 2024

    Pith/arXiv arXiv 2024

  36. [36]

    Fine- tuning vision-language-action models: Optimizing speed and success.arXiv preprint arXiv:2502.19645, 2025

    Moo Jin Kim, Chelsea Finn, and Percy Liang. Fine- tuning vision-language-action models: Optimizing speed and success.arXiv preprint arXiv:2502.19645, 2025

    Pith/arXiv arXiv 2025

  37. [37]

    Room-across-room: Multilingual vision-and-language navigation with dense spatiotempo- ral grounding.arXiv preprint arXiv:2010.07954, 2020

    Alexander Ku, Peter Anderson, Roma Patel, Eugene Ie, and Jason Baldridge. Room-across-room: Multilingual vision-and-language navigation with dense spatiotempo- ral grounding.arXiv preprint arXiv:2010.07954, 2020

    Pith/arXiv arXiv 2010

  38. [38]

    Simplevla-rl: Scaling vla training via reinforcement learning.arXiv preprint arXiv:2509.09674, 2025

    Haozhan Li, Yuxin Zuo, Jiale Yu, Yuhao Zhang, Zhao- hui Yang, Kaiyan Zhang, Xuekai Zhu, Yuchen Zhang, Tianxing Chen, Ganqu Cui, et al. Simplevla-rl: Scaling vla training via reinforcement learning.arXiv preprint arXiv:2509.09674, 2025

    Pith/arXiv arXiv 2025

  39. [39]

    Robogsim: A real2sim2real robotic gaussian splatting simulator.arXiv preprint arXiv:2411.11839, 2024

    Xinhai Li, Jialin Li, Ziheng Zhang, Rui Zhang, Fan Jia, Tiancai Wang, Haoqiang Fan, Kuo-Kun Tseng, and Ruip- ing Wang. Robogsim: A real2sim2real robotic gaussian splatting simulator.arXiv preprint arXiv:2411.11839, 2024

    Pith/arXiv arXiv 2024

  40. [40]

    Gr-rl: Going dexterous and precise for long-horizon robotic manipulation.arXiv preprint arXiv:2512.01801, 2025

    Yunfei Li, Xiao Ma, Jiafeng Xu, Yu Cui, Zhongren Cui, Zhigang Han, Liqun Huang, Tao Kong, Yuxiao Liu, Hao Niu, et al. Gr-rl: Going dexterous and precise for long-horizon robotic manipulation.arXiv preprint arXiv:2512.01801, 2025

    arXiv 2025

  41. [41]

    Flow matching for generative modeling.arXiv preprint arXiv:2210.02747, 2022

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maxim- ilian Nickel, and Matt Le. Flow matching for generative modeling.arXiv preprint arXiv:2210.02747, 2022

    Pith/arXiv arXiv 2022

  42. [42]

    Flow-grpo: Training flow matching models via online rl.arXiv preprint arXiv:2505.05470, 2025

    Jie Liu, Gongye Liu, Jiajun Liang, Yangguang Li, Jia- heng Liu, Xintao Wang, Pengfei Wan, Di Zhang, and Wanli Ouyang. Flow-grpo: Training flow matching models via online rl.arXiv preprint arXiv:2505.05470, 2025

    Pith/arXiv arXiv 2025

  43. [43]

    What can rl bring to vla generalization? an empirical study.arXiv preprint arXiv:2505.19789, 2025

    Jijia Liu, Feng Gao, Bingwen Wei, Xinlei Chen, Qingmin Liao, Yi Wu, Chao Yu, and Yu Wang. What can rl bring to vla generalization? an empirical study.arXiv preprint arXiv:2505.19789, 2025

    arXiv 2025

  44. [44]

    Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017

    Pith/arXiv arXiv 2017

  45. [45]

    Serl: A software suite for sample-efficient robotic reinforcement learning

    Jianlan Luo, Zheyuan Hu, Charles Xu, You Liang Tan, Jacob Berg, Archit Sharma, Stefan Schaal, Chelsea Finn, Abhishek Gupta, and Sergey Levine. Serl: A software suite for sample-efficient robotic reinforcement learning. In2024 IEEE International Conference on Robotics and Automation (ICRA), pages 16961–16969. IEEE, 2024

    2024

  46. [46]

    Sim-and-real co-training: A simple recipe for vision-based robotic manipulation.arXiv preprint arXiv:2503.24361, 2025

    Abhiram Maddukuri, Zhenyu Jiang, Lawrence Yun- liang Chen, Soroush Nasiriany, Yuqi Xie, Yu Fang, Wenqi Huang, Zu Wang, Zhenjia Xu, Nikita Chernyadev, et al. Sim-and-real co-training: A simple recipe for vision-based robotic manipulation.arXiv preprint arXiv:2503.24361, 2025

    Pith/arXiv arXiv 2025

  47. [47]

    Isaac gym: High performance gpu-based physics simulation for robot learning.arXiv preprint arXiv:2108.10470, 2021

    Viktor Makoviychuk, Lukasz Wawrzyniak, Yunrong Guo, Michelle Lu, Kier Storey, Miles Macklin, David Hoeller, Nikita Rudin, Arthur Allshire, Ankur Handa, et al. Isaac gym: High performance gpu-based physics simulation for robot learning.arXiv preprint arXiv:2108.10470, 2021

    Pith/arXiv arXiv 2021

  48. [48]

    Mimicgen: A data generation system for scalable robot learning using human demonstrations

    Ajay Mandlekar, Soroush Nasiriany, Bowen Wen, Ire- tiayo Akinola, Yashraj Narang, Linxi Fan, Yuke Zhu, and Dieter Fox. Mimicgen: A data generation system for scalable robot learning using human demonstrations. arXiv preprint arXiv:2310.17596, 2023

    Pith/arXiv arXiv 2023

  49. [49]

    Active domain randomiza- tion

    Bhairav Mehta, Manfred Diaz, Florian Golemo, Christo- pher J Pal, and Liam Paull. Active domain randomiza- tion. InConference on Robot Learning, pages 1162–

  50. [50]

    Maniskill: Generalizable manipulation skill bench- mark with large-scale demonstrations.arXiv preprint arXiv:2107.14483, 2021

    Tongzhou Mu, Zhan Ling, Fanbo Xiang, Derek Yang, Xuanlin Li, Stone Tao, Zhiao Huang, Zhiwei Jia, and Hao Su. Maniskill: Generalizable manipulation skill bench- mark with large-scale demonstrations.arXiv preprint arXiv:2107.14483, 2021

    Pith/arXiv arXiv 2021

  51. [51]

    Robocasa: Large-scale simulation of everyday tasks for generalist robots.arXiv preprint arXiv:2406.02523, 2024

    Soroush Nasiriany, Abhiram Maddukuri, Lance Zhang, Adeet Parikh, Aaron Lo, Abhishek Joshi, Ajay Man- dlekar, and Yuke Zhu. Robocasa: Large-scale simulation of everyday tasks for generalist robots.arXiv preprint arXiv:2406.02523, 2024

    Pith/arXiv arXiv 2024

  52. [52]

    An algo- rithmic perspective on imitation learning.Foundations and Trends® in Robotics, 7(1-2):1–179, 2018

    Takayuki Osa, Joni Pajarinen, Gerhard Neumann, J An- drew Bagnell, Pieter Abbeel, and Jan Peters. An algo- rithmic perspective on imitation learning.Foundations and Trends® in Robotics, 7(1-2):1–179, 2018

    2018

  53. [53]

    Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0

    Abby O’Neill, Abdul Rehman, Abhiram Maddukuri, Ab- hishek Gupta, Abhishek Padalkar, Abraham Lee, Acorn Pooley, Agrim Gupta, Ajay Mandlekar, Ajinkya Jain, et al. Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0. In2024 IEEE International Conference on Robotics and Automation (ICRA), pages 6892–6903. IEEE, 2024

    2024

  54. [54]

    Sim-to-real transfer of robotic control with dynamics randomization

    Xue Bin Peng, Marcin Andrychowicz, Wojciech Zaremba, and Pieter Abbeel. Sim-to-real transfer of robotic control with dynamics randomization. In2018 IEEE international conference on robotics and automa- tion (ICRA), pages 3803–3810. IEEE, 2018

    2018

  55. [55]

    A reduction of imitation learning and structured prediction to no-regret online learning

    St ´ephane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. InProceedings of the fourteenth international conference on artificial intelli- gence and statistics, pages 627–635. JMLR Workshop and Conference Proceedings, 2011

    2011

  56. [56]

    Habitat: A platform for embodied ai research

    Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied ai research. InProceedings of the IEEE/CVF international conference on computer vision, pages 9339–9347, 2019

    2019

  57. [57]

    Laion-5b: An open large-scale dataset for training next generation image-text models.Advances in neural information processing systems, 35:25278– 25294, 2022

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models.Advances in neural information processing systems, 35:25278– 25294, 2022

    2022

  58. [58]

    Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347, 2017

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347, 2017

    Pith/arXiv arXiv 2017

  59. [59]

    Videovla: Video generators can be generalizable robot manipulators.arXiv preprint arXiv:2512.06963, 2025

    Yichao Shen, Fangyun Wei, Zhiying Du, Yaobo Liang, Yan Lu, Jiaolong Yang, Nanning Zheng, and Baining Guo. Videovla: Video generators can be generalizable robot manipulators.arXiv preprint arXiv:2512.06963, 2025

    arXiv 2025

  60. [60]

    Maniskill3: Gpu parallelized robotics simulation and rendering for gener- alizable embodied ai.arXiv preprint arXiv:2410.00425, 2024

    Stone Tao, Fanbo Xiang, Arth Shukla, Yuzhe Qin, Xander Hinrichsen, Xiaodi Yuan, Chen Bao, Xinsong Lin, Yulin Liu, Tse-kai Chan, et al. Maniskill3: Gpu parallelized robotics simulation and rendering for gener- alizable embodied ai.arXiv preprint arXiv:2410.00425, 2024

    Pith/arXiv arXiv 2024

  61. [61]

    Evaluating gemini robotics policies in a veo world simulator.arXiv preprint arXiv:2512.10675, 2025

    Gemini Robotics Team, Krzysztof Choromanski, Coline Devin, Yilun Du, Debidatta Dwibedi, Ruiqi Gao, Ab- hishek Jindal, Thomas Kipf, Sean Kirmani, Isabel Leal, et al. Evaluating gemini robotics policies in a veo world simulator.arXiv preprint arXiv:2512.10675, 2025

    arXiv 2025

  62. [62]

    Gemma 3 technical report.arXiv preprint arXiv:2503.19786, 2025

    Gemma Team, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Per- rin, Tatiana Matejovicova, Alexandre Ram ´e, Morgane Rivi`ere, et al. Gemma 3 technical report.arXiv preprint arXiv:2503.19786, 2025

    Pith/arXiv arXiv 2025

  63. [63]

    Octo: An open-source generalist robot policy.arXiv preprint arXiv:2405.12213, 2024

    Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joey Hejna, Tobias Kreiman, Charles Xu, et al. Octo: An open-source generalist robot policy.arXiv preprint arXiv:2405.12213, 2024

    Pith/arXiv arXiv 2024

  64. [64]

    Vision-and-dialog navigation

    Jesse Thomason, Michael Murray, Maya Cakmak, and Luke Zettlemoyer. Vision-and-dialog navigation. In Conference on Robot Learning, pages 394–406. PMLR, 2020

    2020

  65. [65]

    Domain ran- domization for transferring deep neural networks from simulation to the real world

    Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. Domain ran- domization for transferring deep neural networks from simulation to the real world. In2017 IEEE/RSJ in- ternational conference on intelligent robots and systems (IROS), pages 23–30. IEEE, 2017

    2017

  66. [66]

    Mujoco: A physics engine for model-based control

    Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control. In2012 IEEE/RSJ international conference on intelligent robots and systems, pages 5026–5033. IEEE, 2012

    2012

  67. [67]

    Rec- onciling reality through simulation: A real-to-sim-to- real approach for robust manipulation, 2024

    Marcel Torne, Anthony Simeonov, Zechu Li, April Chan, Tao Chen, Abhishek Gupta, and Pulkit Agrawal. Rec- onciling reality through simulation: A real-to-sim-to- real approach for robust manipulation, 2024. URL https://arxiv.org/abs/2403.03949

    Pith/arXiv arXiv 2024

  68. [68]

    Llama 2: Open foundation and fine-tuned chat models

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023

    Pith/arXiv arXiv 2023

  69. [69]

    Bridgedata v2: A dataset for robot learning at scale

    Homer Rich Walke, Kevin Black, Tony Z Zhao, Quan Vuong, Chongyi Zheng, Philippe Hansen-Estruch, An- dre Wang He, Vivek Myers, Moo Jin Kim, Max Du, et al. Bridgedata v2: A dataset for robot learning at scale. InConference on Robot Learning, pages 1723–

  70. [70]

    Empirical analysis of sim-and-real cotraining of diffusion poli- cies for planar pushing from pixels.arXiv preprint arXiv:2503.22634, 2025

    Adam Wei, Abhinav Agarwal, Boyuan Chen, Rohan Bosworth, Nicholas Pfaff, and Russ Tedrake. Empirical analysis of sim-and-real cotraining of diffusion poli- cies for planar pushing from pixels.arXiv preprint arXiv:2503.22634, 2025

    Pith/arXiv arXiv 2025

  71. [71]

    Rl-gsbridge: 3d gaussian splatting based real2sim2real method for robotic manipulation learning

    Yuxuan Wu, Lei Pan, Wenhua Wu, Guangming Wang, Yanzi Miao, Fan Xu, and Hesheng Wang. Rl-gsbridge: 3d gaussian splatting based real2sim2real method for robotic manipulation learning. In2025 IEEE International Conference on Robotics and Automation (ICRA), pages 192–198. IEEE, 2025

    2025

  72. [72]

    Qwen3 technical report

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chen- gen Huang, Chenxu Lv, et al. Qwen3 technical report. arXiv preprint arXiv:2505.09388, 2025

    Pith/arXiv arXiv 2025

  73. [73]

    Invari- ance co-training for robot visual generalization.arXiv preprint arXiv:2512.05230, 2025

    Jonathan Yang, Chelsea Finn, and Dorsa Sadigh. Invari- ance co-training for robot visual generalization.arXiv preprint arXiv:2512.05230, 2025

    arXiv 2025

  74. [74]

    Barron, Al- berto Rodriguez, Phillip Isola, and Tsung-Yi Lin

    Lin Yen-Chen, Pete Florence, Jonathan T. Barron, Al- berto Rodriguez, Phillip Isola, and Tsung-Yi Lin. Inerf: Inverting neural radiance fields for pose estimation, 2021. URL https://arxiv.org/abs/2012.05877

    Pith/arXiv arXiv 2021

  75. [75]

    Natural language can help bridge the sim2real gap.arXiv preprint arXiv:2405.10020, 2024

    Albert Yu, Adeline Foote, Raymond Mooney, and Roberto Mart´ın-Mart´ın. Natural language can help bridge the sim2real gap.arXiv preprint arXiv:2405.10020, 2024

    Pith/arXiv arXiv 2024

  76. [76]

    Rlinf: Flexible and efficient large- scale reinforcement learning via macro-to-micro flow transformation.arXiv preprint arXiv:2509.15965, 2025

    Chao Yu, Yuanqing Wang, Zhen Guo, Hao Lin, Si Xu, Hongzhi Zang, Quanlu Zhang, Yongji Wu, Chunyang Zhu, Junhao Hu, et al. Rlinf: Flexible and efficient large- scale reinforcement learning via macro-to-micro flow transformation.arXiv preprint arXiv:2509.15965, 2025

    arXiv 2025

  77. [77]

    Meta- world: A benchmark and evaluation for multi-task and meta reinforcement learning, 2021

    Tianhe Yu, Deirdre Quillen, Zhanpeng He, Ryan Julian, Avnish Narayan, Hayden Shively, Adithya Bellathur, Karol Hausman, Chelsea Finn, and Sergey Levine. Meta- world: A benchmark and evaluation for multi-task and meta reinforcement learning, 2021. URL https://arxiv. org/abs/1910.10897

    Pith/arXiv arXiv 2021

  78. [78]

    Rlinf-vla: A unified and efficient frame- work for reinforcement learning of vision-language- action models, 2026

    Hongzhi Zang, Mingjie Wei, Si Xu, Yongji Wu, Zhen Guo, Yuanqing Wang, Hao Lin, Peihong Wang, Liangzhi Shi, Yuqing Xie, Zhexuan Xu, Zhihao Liu, Kang Chen, Wenhao Tang, Quanlu Zhang, Weinan Zhang, Chao Yu, and Yu Wang. Rlinf-vla: A unified and efficient frame- work for reinforcement learning of vision-language- action models, 2026. URL https://arxiv.org/abs...

    2026

  79. [79]

    Real- to-sim robot policy evaluation with gaussian splatting simulation of soft-body interactions.arXiv preprint arXiv:2511.04665, 2025

    Kaifeng Zhang, Shuo Sha, Hanxiao Jiang, Matthew Loper, Hyunjong Song, Guangyan Cai, Zhuo Xu, Xi- aochen Hu, Changxi Zheng, and Yunzhu Li. Real- to-sim robot policy evaluation with gaussian splatting simulation of soft-body interactions.arXiv preprint arXiv:2511.04665, 2025

    arXiv 2025

  80. [80]

    Vlabench: A large-scale benchmark for language-conditioned robotics manipulation with long-horizon reasoning tasks, 2024

    Shiduo Zhang, Zhe Xu, Peiju Liu, Xiaopeng Yu, Yuan Li, Qinghui Gao, Zhaoye Fei, Zhangyue Yin, Zuxuan Wu, Yu-Gang Jiang, and Xipeng Qiu. Vlabench: A large-scale benchmark for language-conditioned robotics manipulation with long-horizon reasoning tasks, 2024. URL https://arxiv.org/abs/2412.18194

    Pith/arXiv arXiv 2024

Showing first 80 references.