Pith. sign in

REVIEW 4 major objections 5 minor 61 references

XPolicyLab: A Unified Standard and Open Ecosystem for Robot Policy Evaluation and Deployment

T0 review · 4 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read XPolicyLab claims that a minimal four-operation adapter contract lowers the cost of connecting robot policies to evaluation environments from O(NM) to O(N+M), with integration time dropping from over five hours to half an hour.

desk verdict A real systems contribution with a credible O(N+M) adapter standard backed by 42 integrated policies, but the paper's own evidence is thinner than its claims and one protocol detail (the unspecified CALL message) needs tightening before the central claim fully holds. read the letter →

arxiv 2608.09892 v1 pith:DTYCNYRU submitted 2026-08-10 cs.RO

XPolicyLab Community , Tianxing Chen , Yue Chen , Tian Nian , Zijian Cai , Guangyu Chen , Wenwei Lin , Qiwei Liang
show 59 more authors
Peicheng Xiang Kailun Su Zixuan Li Junyuan Tang Yan Qin Qiangyu Chen Shaolong Zhu Xiang Li Jiahao Zhang Weijie Wan Baijun Chen Honghao Su Kehe Ye Shujia Liu Kaixuan Wang Haotian Liang Yunze Liu Mingleyang Li Yuran Wang Boyu Chen Hongzhe Bi Shuhe Huang Hengkai Tan Jisong Cai Yao Mu Jun Guo Xiaofeng Wang Zheng Zhu Weijie Ke Hengtao Li Yuhang Tang Xiaofan Li Ganlin Yang Zhangzheng Tu Shuai Yang Wenxuan Song Pengxiang Ding Kaidong Zhang Yu Sun Junliang Guo Tong Zhang Yixing Chen Rongxu Cui Zongzheng Zhang Haoxiang Ma Junhao Cai Haoyu Zhang Senqiao Yang Jinhui Ye Pengguang Chen Shu Liu Xiu Su Wenhan Fang Wenhao Li Yichao Cao Chengyao Wang Qiang Chen Ping Luo Wenbo Ding
This is my paper · ORCID
classification cs.RO
keywords robotpolicyevaluationadaptercontractobservationschemaactiondependencyisolationclient-serverservingintegrationeffortagent-assistedconformance
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that the dominant cost in robot policy evaluation is pairwise integration: every policy must be hand-wired to every benchmark, simulator, or physical robot, so the effort scales as $O(NM)$ with the number of policies $N$ and environments $M$. It introduces a minimal adapter contract—four operations, `update_obs`, `get_action`, `reset`, plus batched variants—alongside standardized observation and action schemas, so that each policy needs one adapter and each environment needs one client, reducing the cost to $O(N+M)$. The ecosystem ships 42 integrated policies and demonstrates the same adapters running across two simulation benchmarks and a physical-robot evaluation platform. A controlled study reports that conforming to the standard cuts the integration effort of a representative vision-language-action policy from over five hours to about two hours, and to thirty minutes when packaged agent skills execute the conformance procedure. If true, the bottleneck in reproducible comparison shifts from reimplementing glue code to the physical trials themselves.

What carries the argument

The load-bearing object is the policy adapter contract: a minimal interface of model construction, `update_obs`, `get_action`, and `reset`, with `update_obs_batch` and `get_action_batch` for parallel evaluation, paired with the standardized observation and action schemas. The contract confines heterogeneity to the policy side: the serving layer handles transport-level operations such as deserialization and image decoding, while the adapter performs only model-specific transformations, which removes a class of silent faults from mismatched image conventions. A dependency-isolated client/server bridge using WebSocket with MessagePack serialization, retry with cached results, and instance-identity checks lets each side keep its native software stack and run locally or remotely, and that isolation is what makes one adapter per policy reusable across arbitrary environments.

What would settle it

Attempt to integrate a policy that cannot communicate over the four-operation contract—for example one whose inference depends on asynchronous event streams or on internal state that must be injected rather than accumulated through `update_obs`—and check whether the shared, fixed environment-facing loop must be modified to host it; any required change to the loop falsifies the $O(N+M)$ claim. A cheaper observational test is to rerun the agent-assisted integration study with a different coding agent from the one used in the paper and see whether the reported thirty-minute time-to-first-successful-rollout reproduces.

Watch

Extended reading notes

Core claim

The paper's claim is that one shared contract can absorb the full heterogeneity of robot policy runtimes. It specifies a four-operation adapter interface and standardized observation and action schemas, and argues that a single adapter per policy plus a single client per environment makes the interconnection cost linear, $O(N+M)$, rather than the $O(NM)$ of pairwise integrations. An observation is represented as $o_t=\{v_t,q_t,p_t,\ell,m_t\}$ with camera-indexed visuals, joint states, Cartesian poses, a language instruction, and metadata, while actions are embodiment-aware in joint or end-effector space. The environment-facing execution loop stays fixed, and all model-specific preprocessing, action decoding, and state management lives inside the adapter; the paper reports that across its 42 adapters the model-specific code varies by an order of magnitude while the loop stays within a few lines of a fixed reference. The same adapters serve simulation benchmarks and standardized real-robot evaluation without modifying the underlying policy inference code.

Load-bearing premise

The argument rests on the assumption that four operations and the shared observation/action schema are expressive enough for every policy and every environment; if some policy needs an operation outside the contract, or some environment needs policy-specific behavior, the fixed loop breaks and the $O(N+M)$ saving does not hold.

Editorial extensions

If this is right

  • One adapter per policy plus one client per environment means adding a new environment costs one client, not one integration per policy.
  • The controlled study's numbers imply that a researcher without prior exposure can wire a representative VLA policy to a simulator in about two hours by hand, or thirty minutes with the packaged agent skills, versus more than five hours from scratch.
  • Reproducing an already-integrated policy's evaluation takes about ten minutes plus checkpoint download, making closed-loop verification cheap enough to run in routine workflows.
  • Shared observation conventions make cross-policy comparisons interpretable, since differences such as a policy's drop under domain randomization are no longer confounded by hidden preprocessing conventions.
  • The same adapter code drives simulation and physical-robot trials, so policy inference implementations can be frozen across evaluation settings.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper leaves implicit that the $O(N+M)$ guarantee depends on the contract being closed: the real stress test is a policy with a genuinely novel interface (tactile sensing, event cameras, or external memory injection) that does not fit the $o_t=\{v_t,q_t,p_t,\ell,m_t\}$ schema, and no amount of in-family coverage settles that.
  • The thirty-minute agent result is tied to one specific coding-agent stack; whether the packaged skills transfer to other agents is an open, easily testable extension that the paper does not claim.
  • A broader corollary of the interpretability argument is that if the contract becomes widespread, leaderboard comparisons become comparisons of policies rather than preprocessing pipelines, which could change how the field audits reported results.
  • The standard deliberately leaves physical variance and calibration unaddressed; a natural next step, not pursued here, is pairing the adapter contract with standardized fleet-calibration and safety protocols so that real-robot trials are comparable across sites.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. XPolicyLab proposes a standardized adapter contract and serving architecture for robot policy evaluation, aiming to replace O(NM) pairwise policy-environment integrations with O(N+M). The paper defines observation, action, and trajectory schemas; a four-operation adapter interface (update_obs, get_action, reset, plus batched variants); and a WebSocket/MessagePack protocol for dependency-isolated client/server execution. It reports 42 integrated policies, deployments on RoboTwin and RoboDojo leaderboards, and a six-participant study measuring integration effort for connecting pi0.5 to RoboDojo simulation, with median times of >5 h from scratch, ~2 h with XPolicyLab, and ~30 min with agent skills.

Significance. The contribution is a potentially valuable infrastructure artifact: the open repository, 42 policy adapters, and three-way deployment across RoboTwin, RoboDojo sim, and RoboDojo-RealEval provide concrete evidence that a single adapter can be reused across benchmarks. The dependency-isolated serving design with retry caching and instance-identifier checks is a thoughtful engineering contribution. However, the headline O(N+M) claim hinges on the adapter contract being closed, and the paper leaves the CALL protocol message unspecified, which reopens the door to policy-specific environment behavior. The integration-effort study is too small and under-reported to support the abstract's unqualified numbers. The work is significant if the contract can be closed and the evaluation strengthened.

major comments (4)
  1. [Section III-C] The protocol message set in Section III-C includes CALL alongside HELLO, PREPARE_CASE, RESET, INFER, TRIAL_END, HEARTBEAT, and CLOSE, but the paper never defines CALL's semantics or states that it is unused. The O(N+M) claim in the abstract and Section I depends on the four-operation adapter contract being exhaustive: if any adapter uses CALL to request policy-specific behavior (e.g., privileged environment state, reward feedback, simulator parameter changes), the environment client must implement that request, re-introducing pairwise integration. Please specify CALL's payload and semantics, or explicitly state that it is reserved for future use and currently unused by all 42 adapters; otherwise the contract is not closed and the central complexity claim is not secured.
  2. [Section VI-A, Table V] The controlled study reports medians over N=6 participants with no variance, no confidence intervals, and no significance testing, and the paper itself describes it as 'an indicative case study rather than a controlled measurement.' Despite this, the abstract and conclusion present the numbers ('from over five hours to two hours' and 'to thirty minutes') as unqualified findings. Please either report per-participant times and appropriate statistics, or explicitly qualify these numbers throughout the paper as an exploratory case study. As written, the quantitative headline claims exceed what the evidence supports.
  3. [Section VI-A] The agent-skills condition differs from the manual XPolicyLab condition by both the presence of a coding agent and the presence of the packaged skills, so the observed reduction from ~2 h to ~30 min cannot be attributed to the skills alone; an agent without the skills (or with generic instructions) would be needed as a control. The sentence 'the comparison is about the skills rather than about model choice' only fixes the agent stack across participants; it does not isolate the skills' contribution. Please add a control condition or soften the claim that the skills themselves are responsible for the order-of-magnitude reduction.
  4. [Abstract, Section I, Section VII] The paper repeatedly claims that 'model-specific code varies by an order of magnitude while the environment-facing loop stays within a few lines of a fixed reference,' but no measurement of adapter code size or environment-loop code size is presented anywhere in the evaluation. Please add a quantitative analysis (e.g., LoC distribution across the 42 adapters and the environment client implementation) or remove the claim, as it is currently unsupported.
minor comments (5)
  1. [Section VI-A] The phrase 'Six participants admit exactly the six orderings' should be 'span' or 'cover' rather than 'admit.'
  2. [Table I] Several policy names have inconsistent spacing (e.g., 'FastW AM', 'LingBot-V A') and the entry 'A1' is ambiguous without a version or reference number.
  3. [Section III-D] The statement 'All fields are optional except those required by the evaluated policy' is circular; please clarify which fields, if any, are mandatory in the standardized schema.
  4. [Section V-C] The sentence 'Each policy is evaluated over 10 trials per task, yielding 180 physical trials across the three embodiments' is ambiguous about whether the 180 trials are per embodiment or in total across all embodiments.
  5. [Section VI-A] The metric 'time-to-first-successful-rollout' is not formally defined; please state what constitutes a successful rollout and when the clock starts and stops.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the O(N+M) claim is a design consequence of the standard, and the effort reductions are measured, not fitted or derived from self-citations.

full rationale

The paper's central claim is not a first-principles derivation. XPolicyLab defines a four-operation adapter contract and a client/server protocol, and the O(N+M) integration bound follows directly from the architecture: one adapter per policy and one environment client per environment. This is a system-design claim whose value lies in the artifact (public repository, 42 adapters), not in a fitted parameter renamed as a prediction. The integration-effort numbers in Table V come from a within-subject study with N=6 and are explicitly described as an indicative case study; no fitted quantity is reused as evidence. The deployment evidence on RoboTwin and RoboDojo does cite benchmarks whose core authors overlap with XPolicyLab, but those benchmarks are public and their leaderboards are externally checkable, and the effort-reduction claim does not rest on those leaderboard numbers. The paper itself flags the main limitations: the contract's breadth is treated as an empirical question (Section IV) and the user study is called indicative rather than controlled (Section VI-A). The unspecified CALL protocol message is a completeness gap in the O(N+M) guarantee, but it is not a circular reduction: nothing in the paper defines CALL in terms of the conclusion or uses it to fit a prediction. Overall, no step in the paper's argument is equivalent to its input by construction.

Assumptions & free parameters 0 free parameters · 3 assumptions · 2 invented entities

The central claim rests on the expressiveness of the four-operation adapter contract and on the standardized schemas being sufficient for all policies and environments. It also assumes that policy-server and environment-client communication over WebSocket/MessagePack is reliable enough for closed-loop control. None of these are proven; they are domain assumptions supported by the 42 adapters and three demonstrated environments. There are no fitted free parameters. The XPolicyLab standard and the agent skills are software artifacts with independent evidence in the public repository and leaderboards.

assumptions (3)
  • domain assumption Every robot policy of interest can be expressed through the four adapter operations (update_obs, get_action, reset, batched variants) and the standardized observation/action schemas.
    The whole O(N+M) claim depends on this expressiveness; stated in Section III-B and tested on 42 policies only.
  • domain assumption Standardized observation and action schemas are sufficient to capture the interactions of all target environments (simulators and physical robots).
    Schemas in Section III-D; only three environments demonstrated (RoboTwin, RoboDojo sim and real).
  • domain assumption Policy server and environment client communicate over WebSocket/MessagePack reliably enough for closed-loop control.
    Protocol in Section III-C; reliability is claimed without formal verification.
invented entities (2)
  • XPolicyLab standard and adapter contract independent evidence
    purpose: Defines a common interface between robot policies and evaluation environments to replace pairwise integrations.
    Public repository, 42 shipped adapters, and public leaderboards provide falsifiable handles outside this paper.
  • Agent skills (.agents/skills/) independent evidence
    purpose: Machine-readable conformance procedures that coding agents execute to scaffold and audit adapters.
    Shipped in the repository; the controlled study's agent condition tests them.

how reviews work

0 comments
Cite this review

Pith. "Pith review of XPolicyLab: A Unified Standard and Open Ecosystem for Robot Policy Evaluation and Deployment." pith.science (2026). https://pith.science/paper/DTYCNYRU

@misc{pith2026260809892,
  author       = {Pith},
  title        = {Pith review of: XPolicyLab: A Unified Standard and Open Ecosystem for Robot Policy Evaluation and Deployment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DTYCNYRU}},
  note         = {Machine review of arXiv:2608.09892}
}
read the original abstract

Robot policy evaluation and deployment remain fragmented by model-specific software dependencies, data representations, and runtime interfaces, so that connecting N policies to M evaluation environments requires O(NM) separate integrations. We present XPolicyLab, a unified standard and open ecosystem that reduces this cost to O(N+M). XPolicyLab specifies common observation, action, and trajectory schemas together with a minimal adapter interface for observation updates, action prediction, batched execution, and episode reset, while a dependency-isolated client/server architecture separates policy inference from environment execution, so that each side retains its native software stack and may run locally or remotely. The ecosystem integrates 42 robot policies and standardizes their installation, debugging, serving, and evaluation workflows. Across these adapters, model-specific code varies by an order of magnitude while the environment-facing loop stays within a few lines of a fixed reference, confirming that the contract confines heterogeneity to the policy side. In a controlled study, conforming to the standard reduces the integration effort of a representative policy from over five hours to two hours, and packaged agent skills reduce it further to thirty minutes. The same adapters serve RoboTwin, RoboDojo simulation, and standardized real-robot evaluation through one interface. XPolicyLab is released as shared infrastructure for reproducible policy comparison and standardized deployment across simulation and physical platforms. Project website: https://xpolicylab.github.io/.

Figures

Figures reproduced from arXiv: 2608.09892 by the authors.

Figure 1
Figure 1. Overview of XPolicyLab. Heterogeneous policy runtimes (left) keep their native dependencies, checkpoints, and processing pipelines, while evaluation environments (right) keep their simulator and robot stacks. XPolicyLab (center) joins the two through a unified lifecycle from installation to evaluation, a minimal adapter contract, standardized observation and action schemas, and a dependency-isolated serving bridge. … view at source ↗
Figure 2
Figure 2. Cross-platform policy evaluation through XPolicyLab. A shared policy codebase and a standardized serving interface allow the same policy integration to be evaluated in RoboTwin 2.0, RoboDojo simulation, and RoboDojo real-world settings with minimal policy-side adaptation. embodiments, and evaluation protocols, yet share the same policy abstraction and serving architecture. The leaderboards therefore characterize the… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

61 extracted references · 16 canonical work pages

  1. [1]

    Openvla: An open- source vision-language-action model,

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. Foster, G. Lam, P. Sanketiet al., “Openvla: An open- source vision-language-action model,”arXiv preprint arXiv:2406.09246, 2024

    arXiv 2024

  2. [2]

    pi0.5: A vision-language-action model with open-world generalization,

    P. Intelligence, K. Black, N. Brown, J. Darpinian, K. Dhabalia, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusaiet al., “ pi0.5: A vision-language-action model with open-world generalization,” arXiv preprint arXiv:2504.16054, 2025

    arXiv 2025

  3. [3]

    Pi 0.7: a steerable generalist robotic foundation model with emergent capabilities,

    P. Intelligence, B. Ai, A. Amin, R. Aniceto, A. Balakrishna, G. Balke, K. Black, G. Bokinsky, S. Cao, T. Charbonnieret al., “Pi 0.7: a steerable generalist robotic foundation model with emergent capabilities,”arXiv preprint arXiv:2604.15483, 2026

    arXiv 2026

  4. [4]

    Octo: An open-source generalist robot policy,

    O. M. Team, D. Ghosh, H. Walke, K. Pertsch, K. Black, O. Mees, S. Dasari, J. Hejna, T. Kreiman, C. Xuet al., “Octo: An open-source generalist robot policy,”arXiv preprint arXiv:2405.12213, 2024

    arXiv 2024

  5. [5]

    World action models are zero-shot policies,

    S. Ye, Y . Ge, K. Zheng, S. Gao, S. Yu, G. Kurian, S. Indupuru, Y . L. Tan, C. Zhu, J. Xianget al., “World action models are zero-shot policies,”arXiv preprint arXiv:2602.15922, 2026

    arXiv 2026

  6. [6]

    Xiaomi-robotics-0: An open-sourced vision-language-action model with real-time execution,

    R. Cai, J. Guo, X. He, P. Jin, J. Li, B. Lin, F. Liu, W. Liu, F. Ma, K. Ma et al., “Xiaomi-robotics-0: An open-sourced vision-language-action model with real-time execution,”arXiv preprint arXiv:2602.12684, 2026

    arXiv 2026

  7. [7]

    Hy-embodied-0.5-vla: From vision- language-action models to a real-world robot learning stack,

    H. Zhang, L. Xiang, H. Lin, Z. Huang, M. Wang, D. Zhong, Y . Dong, Y . Wu, Y . Rao, D. Zhanget al., “Hy-embodied-0.5-vla: From vision- language-action models to a real-world robot learning stack,”arXiv preprint arXiv:2606.14409, 2026

    arXiv 2026

  8. [8]

    G3flow: Generative 3d semantic flow for pose-aware and generalizable object manipulation,

    T. Chen, Y . Mu, Z. Liang, Z. Chen, S. Peng, Q. Chen, M. Xu, R. Hu, H. Zhang, X. Liet al., “G3flow: Generative 3d semantic flow for pose-aware and generalizable object manipulation,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 1735–1744

    work page 2025

Show all 61 references
  1. [9]

    Towards trustworthy embodied intelligence: A systems framework and graded trustworthiness levels,

    X. Yang, T. Chen, H. Su, M. Wang, C. Yu, Z. Tu, Y . Chen, Y . Huo, L. Zhang, Y . Huanget al., “Towards trustworthy embodied intelligence: A systems framework and graded trustworthiness levels,”arXiv preprint arXiv:2607.26121, 2026

    2026 arXiv

  2. [10]

    Lerobot: An open-source library for end-to-end robot learning,

    R. Cadene, S. Alibert, F. Capuano, M. Aractingi, A. Zouitine, P. Kooijmans, J. Choghari, M. Russi, C. Pascal, S. Palmaet al., “Lerobot: An open-source library for end-to-end robot learning,” in International Conference on Learning Representations, vol. 2026, 2026, pp. 122 398–122 417

    2026

  3. [11]

    Fastumi: A scalable and hardware- independent universal manipulation interface with dataset,

    K. Liu, C. Guan, Z. Jia, Z. Wu, X. Liu, T. Wang, S. Liang, P. Chen, P. Zhang, H. Songet al., “Fastumi: A scalable and hardware- independent universal manipulation interface with dataset,”arXiv preprint arXiv:2409.19499, 2024

    2024 arXiv

  4. [12]

    Droid: A large-scale in-the-wild robot manipulation dataset,

    A. Khazatsky, K. Pertsch, S. Nair, A. Balakrishna, S. Dasari, S. Karam- cheti, S. Nasiriany, M. K. Srirama, L. Y . Chen, K. Elliset al., “Droid: A large-scale in-the-wild robot manipulation dataset,”arXiv preprint arXiv:2403.12945, 2024

    2024 arXiv

  5. [13]

    Starvla: A lego-like codebase for vision-language- action model developing,

    S. Community, “Starvla: A lego-like codebase for vision-language- action model developing,”arXiv preprint arXiv:2604.05014, 2026

    2026 arXiv

  6. [14]

    {RLinf}: Flexible and efficient {Large-Scale} reinforcement learning via {Macro-to-Micro} flow transformation,

    C. Yu, Y . Wang, Z. Guo, H. Lin, S. Xu, H. Zang, Q. Zhang, Y . Wu, C. Zhu, J. Huet al., “ {RLinf}: Flexible and efficient {Large-Scale} reinforcement learning via {Macro-to-Micro} flow transformation,” in20th USENIX Symposium on Operating Systems Design and Implementation (OSD...

    2026

  7. [15]

    Robotwin 2.0: A scalable data generator and benchmark with strong domain randomization for robust bimanual robotic manipulation,

    T. Chen, Z. Chen, B. Chen, Z. Cai, Y . Liu, Z. Li, Q. Liang, X. Lin, Y . Ge, Z. Guet al., “Robotwin 2.0: A scalable data generator and benchmark with strong domain randomization for robust bimanual robotic manipulation,”arXiv preprint arXiv:2506.18088, 2025

    2025 arXiv

  8. [16]

    RoboDojo: A unified sim-and-real benchmark for comprehensive evaluation of generalist robot manipulation policies,

    T. Chen, Y . Chen, Z. Li, J. Tang, K. Su, W. Wan, B. Chen, H. Lu, H. Yan, H. Suet al., “RoboDojo: A unified sim-and-real benchmark for comprehensive evaluation of generalist robot manipulation policies,” arXiv preprint arXiv:2607.04434, 2026

    2026 arXiv

  9. [17]

    Bridge data: Boosting gener- alization of robotic skills with cross-domain datasets,

    F. Ebert, Y . Yang, K. Schmeckpeper, B. Bucher, G. Georgakis, K. Daniilidis, C. Finn, and S. Levine, “Bridge data: Boosting gener- alization of robotic skills with cross-domain datasets,”arXiv preprint arXiv:2109.13396, 2021

    2021 arXiv

  10. [18]

    Rlbench: The robot learning benchmark & learning environment,

    S. James, Z. Ma, D. R. Arrojo, and A. J. Davison, “Rlbench: The robot learning benchmark & learning environment,” 2019. [Online]. Available: https://arxiv.org/abs/1909.12271

    2019 arXiv

  11. [19]

    Maniskill2: A unified benchmark for generalizable manipulation skills,

    J. Gu, F. Xiang, X. Li, Z. Ling, X. Liu, T. Mu, Y . Tang, S. Tao, X. Wei, Y . Yaoet al., “Maniskill2: A unified benchmark for generalizable manipulation skills,” inThe Eleventh International Conference on Learning Representations, 2023

    2023

  12. [20]

    Libero: Benchmarking knowledge transfer for lifelong robot learning,

    B. Liu, Y . Zhu, C. Gao, Y . Feng, Q. Liu, Y . Zhu, and P. Stone, “Libero: Benchmarking knowledge transfer for lifelong robot learning,” Advances in Neural Information Processing Systems, vol. 36, pp. 44 776– 44 791, 2023

    2023

  13. [21]

    Rmbench: Memory-dependent robotic manipulation benchmark with insights into policy design,

    T. Chen, Y . Wang, M. Li, Y . Qin, H. Shi, Z. Li, Y . Hu, Y . Zhang, K. Wang, Y . Chenet al., “Rmbench: Memory-dependent robotic manipulation benchmark with insights into policy design,”arXiv preprint arXiv:2603.01229, 2026

    2026 arXiv

  14. [22]

    Univtac: A unified simulation platform for visuo- tactile manipulation data generation, learning, and benchmarking,

    B. Chen, W. Wan, T. Chen, X. Guo, C. Xu, Y . Qi, H. Zhang, L. Wu, T. Xu, Z. Liet al., “Univtac: A unified simulation platform for visuo- tactile manipulation data generation, learning, and benchmarking,” arXiv preprint arXiv:2602.10093, 2026

    2026

  15. [23]

    Robotwin: Dual-arm robot benchmark with generative digital twins,

    Y . Mu, T. Chen, Z. Chen, S. Peng, Z. Lan, Z. Gao, Z. Liang, Q. Yu, Y . Zou, M. Xuet al., “Robotwin: Dual-arm robot benchmark with generative digital twins,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 27 649–27 660

    2025

  16. [24]

    Benchmarking generalizable bimanual manipulation: Robotwin dual-arm collaboration challenge at cvpr 2025 meis workshop,

    T. Chen, K. Wang, Z. Yang, Y . Zhang, Z. Chen, B. Chen, W. Dong, Z. Liu, D. Chen, T. Yanget al., “Benchmarking generalizable bimanual manipulation: Robotwin dual-arm collaboration challenge at cvpr 2025 meis workshop,”arXiv preprint arXiv:2506.23351, 2025

    2025 arXiv

  17. [25]

    Roboarena: Distributed real-world evaluation of generalist robot policies,

    P. Atreya, K. Pertsch, T. Lee, M. J. Kim, A. Jain, A. Kuramshin, C. Eppner, C. Neary, E. Hu, F. Ramoset al., “Roboarena: Distributed real-world evaluation of generalist robot policies,”arXiv preprint arXiv:2506.18123, 2025

    2025

  18. [26]

    Robochallenge: Large-scale real-robot evaluation of embodied policies,

    A. Yakefu, B. Xie, C. Xu, E. Zhang, E. Zhou, F. Jia, H. Yang, H. Fan, H. Zhang, H. Penget al., “Robochallenge: Large-scale real-robot evaluation of embodied policies,”arXiv preprint arXiv:2510.17950, 2025

    2025

  19. [27]

    pi0: A vision-language-action flow model for general robot control,

    K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichteret al., “ pi0: A vision-language-action flow model for general robot control,”arXiv preprint arXiv:2410.24164, 2024

    2024 arXiv

  20. [28]

    Gr00t n1: An open foundation model for generalist humanoid robots,

    J. Bjorck, F. Casta˜neda, N. Cherniadev, X. Da, R. Ding, L. Fan, Y . Fang, D. Fox, F. Hu, S. Huanget al., “Gr00t n1: An open foundation model for generalist humanoid robots,”arXiv preprint arXiv:2503.14734, 2025

    2025 arXiv

  21. [29]

    A1: A fully transparent open-source, adaptive and efficient truncated vision-language-action model,

    K. Zhang, J. Zhang, R. Xu, Y . Sun, S. Xue, Y . Wen, X. Guo, M. Guo, W. Liufu, L. Zihouet al., “A1: A fully transparent open-source, adaptive and efficient truncated vision-language-action model,”arXiv preprint arXiv:2604.05672, 2026

    2026 arXiv

  22. [30]

    Aha-wam: Asynchronous horizon-adaptive world-action modeling with observation-guided context routing,

    J. Cai, L. Ling, S. Chu, Z. Liu, J. Kang, Z. Liang, W. Xu, Y . Mao, W. Zhang, X. Yanget al., “Aha-wam: Asynchronous horizon-adaptive world-action modeling with observation-guided context routing,”arXiv preprint arXiv:2606.09811, 2026

    2026 arXiv

  23. [31]

    Abot-m0: Vla foundation model for robotic manipulation with action manifold learning,

    Y . Yang, S. Zeng, T. Lin, X. Chang, D. Qi, J. Xiao, H. Liu, R. Chen, Y . Chen, D. Huoet al., “Abot-m0: Vla foundation model for robotic manipulation with action manifold learning,”arXiv preprint arXiv:2602.11236, 2026

    2026 arXiv

  24. [32]

    Learning fine- grained bimanual manipulation with low-cost hardware,

    T. Z. Zhao, V . Kumar, S. Levine, and C. Finn, “Learning fine- grained bimanual manipulation with low-cost hardware,”arXiv preprint arXiv:2304.13705, 2023

    2023 arXiv

  25. [33]

    Being-h0. 5: Scaling human-centric robot learning for cross-embodiment generalization,

    H. Luo, Y . Wang, W. Zhang, S. Zheng, Z. Xi, C. Xu, H. Xu, H. Yuan, C. Zhang, Y . Wanget al., “Being-h0. 5: Scaling human-centric robot learning for cross-embodiment generalization,”arXiv preprint arXiv:2601.12993, 2026

    2026

  26. [34]

    Dm0: An embodied-native vision-language- action model towards physical ai,

    E. Yu, H. Lv, J. Sun, K. Lin, R. Zhang, Y . Shi, Y . Chen, Z. Chen, Z. Zhang, F. Jiaet al., “Dm0: An embodied-native vision-language- action model towards physical ai,”arXiv preprint arXiv:2602.14974, 2026

    2026

  27. [35]

    Dexora: Open-source vla for high-dof bimanual dexterity,

    Z. Zhang, J. Pang, Z. Yang, K. Li, M. Liao, S. Zhang, G. Chi, J. Guo, H.- a. Gao, M. Shiet al., “Dexora: Open-source vla for high-dof bimanual dexterity,”arXiv preprint arXiv:2605.18722, 2026

    2026 arXiv

  28. [36]

    Diffusion policy: Visuomotor policy learning via action diffusion,

    C. Chi, Z. Xu, S. Feng, E. Cousineau, Y . Du, B. Burchfiel, R. Tedrake, and S. Song, “Diffusion policy: Visuomotor policy learning via action diffusion,”The International Journal of Robotics Research, vol. 44, no. 10-11, pp. 1684–1704, 2025

    2025

  29. [37]

    Eventvla: Event-driven visual evidence memory for long-horizon vision-language-action policies,

    G. Yang, Z. Tu, Y . Yang, S. Mao, J. Dong, T. Chen, J. Peng, J. Xiong, J. Cao, J. Daiet al., “Eventvla: Event-driven visual evidence memory for long-horizon vision-language-action policies,” arXiv preprint arXiv:2606.20092, 2026

    2026 arXiv

  30. [38]

    Fast-wam: Do world action models need test-time future imagination?

    T. Yuan, Z. Dong, Y . Liu, and H. Zhao, “Fast-wam: Do world action models need test-time future imagination?”arXiv preprint arXiv:2603.16666, 2026

    2026 arXiv

  31. [39]

    Galaxea g0.5 technical report,

    Galaxea Team, “Galaxea g0.5 technical report,” 2026. [Online]. Available: https://opengalaxea.github.io/G05/

    2026

  32. [40]

    Galaxea open-world dataset and g0 dual-system vla model,

    T. Jiang, T. Yuan, Y . Liu, C. Lu, J. Cui, X. Liu, S. Cheng, J. Gao, H. Xu, and H. Zhao, “Galaxea open-world dataset and g0 dual-system vla model,”arXiv preprint arXiv:2509.00576, 2025

    2025 arXiv

  33. [41]

    Gigaworld-policy: An efficient action-centered world–action model,

    A. Ye, B. Wang, C. Ni, G. Huang, G. Zhao, H. Li, H. Li, J. Li, J. Lv, J. Liuet al., “Gigaworld-policy: An efficient action-centered world–action model,”arXiv preprint arXiv:2603.17240, 2026

    2026

  34. [42]

    Agibot world colosseo: A large-scale manipulation platform for scalable and intelligent embodied systems,

    Q. Bu, J. Cai, L. Chen, X. Cui, Y . Ding, S. Feng, S. Gao, X. He, X. Hu, X. Huanget al., “Agibot world colosseo: A large-scale manipulation platform for scalable and intelligent embodied systems,”arXiv preprint arXiv:2503.06669, 2025

    2025 arXiv

  35. [43]

    H-rdt: Human manipulation enhanced bimanual robotic manipulation,

    H. Bi, L. Wu, T. Lin, H. Tan, Z. Su, H. Su, and J. Zhu, “H-rdt: Human manipulation enhanced bimanual robotic manipulation,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 40, no. 22, 2026, pp. 18 135–18 143

    2026

  36. [44]

    Internvla-a1: Unifying understanding, generation and action for robotic manipulation,

    J. Cai, Z. Cai, J. Cao, Y . Chen, Z. He, L. Jiang, H. Li, H. Li, Y . Li, Y . Liuet al., “Internvla-a1: Unifying understanding, generation and action for robotic manipulation,”arXiv preprint arXiv:2601.02456, 2026

    2026

  37. [45]

    Internvla-a1. 5: Unifying understanding, latent foresight, and action for compositional generalization,

    H. Ma, J. Cai, X. Xu, H. Li, Y . Yang, Y . Tian, J. Cao, H. Zhu, Z. Qiu, Y . Yanget al., “Internvla-a1. 5: Unifying understanding, latent foresight, and action for compositional generalization,”arXiv preprint arXiv:2607.04988, 2026

    2026 arXiv

  38. [46]

    Lda-1b: Scaling latent dynamics action model via universal embodied data ingestion,

    J. Lyu, K. Liu, X. Zhang, H. Liao, Y . Feng, W. Zhu, T. Shen, J. Chen, J. Zhang, Y . Donget al., “Lda-1b: Scaling latent dynamics action model via universal embodied data ingestion,”arXiv preprint arXiv:2602.12215, 2026

    2026 arXiv

  39. [47]

    Causal world modeling for robot control,

    L. Li, Q. Zhang, Y . Luo, S. Yang, R. Wang, F. Han, M. Yu, Z. Gao, N. Xue, X. Zhu, Y . Shen, and Y . Xu, “Causal world modeling for robot control,” 2026. [Online]. Available: https://arxiv.org/abs/2601.21998

    2026 arXiv

  40. [48]

    A pragmatic vla foundation model,

    W. Wu, F. Lu, Y . Wang, S. Yang, S. Liu, F. Wang, Q. Zhu, H. Sun, Y . Wang, S. Maet al., “A pragmatic vla foundation model,”arXiv preprint arXiv:2601.18692, 2026

    2026 arXiv

  41. [49]

    Molmoact2: Action reasoning models for real-world deployment,

    H. Fang, J. Duan, D. Clay, S. Wang, S. Liu, W. Huang, X. Fan, W.-C. Tsai, S. Chen, Y . R. Wanget al., “Molmoact2: Action reasoning models for real-world deployment,”arXiv preprint arXiv:2605.02881, 2026

    2026 arXiv

  42. [50]

    Starvla- α: Reducing complexity in vision-language- action systems,

    J. Ye, N. Gao, S. Yang, J. Zheng, Z. Wang, Y . Chen, P. Chen, Y . Chen, S. Liu, and J. Jia, “Starvla- α: Reducing complexity in vision-language- action systems,”arXiv preprint arXiv:2604.11757, 2026

    2026 arXiv

  43. [51]

    Rdt-1b: a diffusion foundation model for bimanual manipula- tion,

    S. Liu, L. Wu, B. Li, H. Tan, H. Chen, Z. Wang, K. Xu, H. Su, and J. Zhu, “Rdt-1b: a diffusion foundation model for bimanual manipula- tion,” inInternational Conference on Learning Representations, vol. 2025, 2025, pp. 29 982–30 009

    2025

  44. [52]

    Rise: Self-improving robot policy with compositional world model,

    J. Yang, K. Lin, J. Li, W. Zhang, T. Lin, L. Wu, Z. Su, H. Zhao, Y .-Q. Zhang, L. Chenet al., “Rise: Self-improving robot policy with compositional world model,”arXiv preprint arXiv:2602.11075, 2026

    2026 arXiv

  45. [53]

    Smolvla: A vision-language-action model for affordable and efficient robotics,

    M. Shukor, D. Aubakirova, F. Capuano, P. Kooijmans, S. Palma, A. Zouitine, M. Aractingi, C. Pascal, M. Russi, A. Marafiotiet al., “Smolvla: A vision-language-action model for affordable and efficient robotics,”arXiv preprint arXiv:2506.01844, 2025

    2025 arXiv

  46. [54]

    Spatial forcing: Implicit spatial representation alignment for vision-language-action model,

    F. Li, W. Song, H. Zhao, J. Wang, P. Ding, D. Wang, L. Zeng, and H. Li, “Spatial forcing: Implicit spatial representation alignment for vision-language-action model,” inInternational Conference on Learning Representations, vol. 2026, 2026, pp. 132 324–132 345

    2026

  47. [55]

    Spirit-v1.5: Clean data is the enemy of great robot foundation models,

    S. A. Team, “Spirit-v1.5: Clean data is the enemy of great robot foundation models,”Spirit AI Blog, 2026, https://www.spirit- ai.com/en/blog/spirit-v1-5

    2026

  48. [56]

    Tinyvla: Towards fast, data-efficient vision- language-action models for robotic manipulation,

    J. Wen, Y . Zhu, J. Li, M. Zhu, Z. Tang, K. Wu, Z. Xu, N. Liu, R. Cheng, C. Shenet al., “Tinyvla: Towards fast, data-efficient vision- language-action models for robotic manipulation,”IEEE Robotics and Automation Letters, 2025

    2025

  49. [57]

    X-vla: Soft-prompted transformer as scalable cross-embodiment vision-language-action model,

    J. Zheng, J. Li, Z. Wang, D. Liu, X. Kang, Y . Feng, Y . Zheng, J. Zou, Y . Chen, J. Zenget al., “X-vla: Soft-prompted transformer as scalable cross-embodiment vision-language-action model,” inInternational Conference on Learning Representations, vol. 2026, 2026, pp. 60 580– 60 606

    2026

  50. [58]

    Unified 4d world action modeling from video priors with asynchronous denoising,

    J. Guo, Q. Li, P. Li, Z. Chen, N. Sun, Y . Su, H. Wang, Y . Zhang, X. Li, and H. Liu, “Unified 4d world action modeling from video priors with asynchronous denoising,”arXiv preprint arXiv:2604.26694, 2026

    2026 arXiv

  51. [59]

    Xiaomi-robotics-1: Scaling vision-language- action models with over 100k hours of real-world trajectories,

    X. R. Team, J. Guo, P. Jin, J. Li, P. Li, Y . Li, F. Liu, W. Peng, O. Qin, Y . Suet al., “Xiaomi-robotics-1: Scaling vision-language- action models with over 100k hours of real-world trajectories,”arXiv preprint arXiv:2607.15330, 2026

    2026 arXiv

  52. [60]

    Unit: Toward a unified physical language for human-to-humanoid policy learning and world modeling,

    T. U. o. H. K. XPENG Robotics, Tsinghua University, “Unit: Toward a unified physical language for human-to-humanoid policy learning and world modeling,”https://xpeng-robotics.github.io/unit/, 2026

    2026

  53. [61]

    Rxbrain: Embodied cognition foundation model with joint language-visual reasoning and imagination,

    H. Liang, M. Chen, Y . Huang, Y . Guo, X. Zhu, X. Shi, K. Wang, Y . Mao, W. Zhou, L. Chenet al., “Rxbrain: Embodied cognition foundation model with joint language-visual reasoning and imagination,”arXiv preprint arXiv:2607.14187, 2026

    2026 arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.