Pith. sign in

REVIEW 4 major objections 5 minor 42 references

AnyBody: A Benchmark Suite for Cross-Embodiment Manipulation

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Multi-embodiment reinforcement learning matches single-robot baselines in-distribution but fails zero-shot on extrapolation and composition, according to a new benchmark for manipulation.

desk verdict A genuinely useful benchmark for cross-embodiment manipulation, with honest but mostly negative results; the main weaknesses are missing seed variance and an over-relied-on cosine similarity analysis that the paper doesn't actually need to carry its argument. read the letter →

arxiv 2505.14986 v1 pith:ICGD2KMQ submitted 2025-05-21 cs.RO cs.LG

Meenal Parakh , Alexandre Kirchmeyer , Beining Han , Jia Deng This is my paper
classification cs.ROcs.LG
keywords cross-embodimentmanipulationmorphologygeneralizationrobotbenchmarkreinforcementlearningzero-shotmulti-embodimentreachtaskpush
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 introduces AnyBody, a simulated benchmark for testing whether a single reinforcement-learning policy can control many robot morphologies and generalize to morphologies it never saw during training. The benchmark organizes generalization into three regimes of increasing difficulty—interpolation within a morphology family, composition of known body parts into a new arrangement, and extrapolation to a structurally different robot—and evaluates two manipulation tasks, reach and push. The central empirical claim is that multi-embodiment agents match single-embodiment baselines on in-distribution training morphologies but largely fail at zero-shot generalization on composition and extrapolation, with transformer-based agents scoring near zero on some push tasks. The paper argues this shows the field still lacks methods that reason over morphology itself.

What carries the argument

The central mechanism is a link-sequence representation of a robot: each link is encoded as a 48-dimensional vector combining geometry, joint type, joint axis, joint origin, and a sinusoidal encoding of the joint position. A cosine similarity between these vectors for train and test morphologies is used to justify the interpolation–composition–extrapolation ordering. Policies are trained with morphology-conditioned PPO, using either an MLP or a transformer that consumes link tokens, with the addition of symlog return normalization, slow critic updates, and discrete action bins for the reach task.

What would settle it

For each AnyBody task, compute the cosine similarity between train and test morphologies and correlate it with the measured zero-shot success rate across multiple methods; if a test robot with higher cosine similarity performs worse than one with lower similarity, or if the ordering interpolation < composition < extrapolation does not consistently appear in the results, the benchmark's central organizing assumption is falsified.

Watch

Extended reading notes

Core claim

AnyBody provides a standardized evaluation for cross-embodiment manipulation with 18 robot variations, two tasks, and a three-axis difficulty split. The authors show that morphology-conditioned multi-embodiment agents, when trained with the same compute budget as single-embodiment baselines, achieve comparable or better in-distribution multi-task scores—especially with a transformer backbone—but their zero-shot performance on unseen morphologies is dramatically lower in composition and extrapolation settings. For instance, in the Arms push task, multi-embodiment agents essentially fail with 0% success rate, and fine-tuning for 30k steps still leaves a significant gap relative to single-embodiment learning. The paper interprets this as evidence that current multi-embodiment training does not yet yield morphological reasoning.

Load-bearing premise

The benchmark's three difficulty levels assume that cosine similarity between hand-crafted 48-dimensional link vectors measures how hard it is for a policy to transfer to a new robot, so that lower similarity means harder generalization.

Editorial extensions

If this is right

  • AnyBody gives the community a common testbed: any cross-embodiment manipulation method can be measured on the same train/test splits and the same three difficulty axes.
  • Simple multi-embodiment training is not enough for zero-shot generalization to structurally novel robots, so methods that explicitly reason about morphology or use part-level composition are needed.
  • Fine-tuning a multi-embodiment checkpoint is faster than training from scratch but still does not close the gap to single-embodiment performance, suggesting that pretrained control representations do not yet encode reusable morphology knowledge.
  • The design choices shown to matter for the reach task—symlog returns, discrete actions, and transformer backbones—provide concrete baselines for future method development.
  • The benchmark's modular codebase supports adding new tasks, obstacle variations, and observation types, so the methodology can be extended beyond reach and push.

Reading between the lines

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

  • The failure on composition tasks suggests that part-level grounded reasoning—understanding what a gripper, plane, or stick contributes to a task—is absent from current end-to-end RL agents, and this capability may need to be explicitly induced.
  • The cosine-similarity ordering could be used as a curriculum signal: training on morphologies with progressively lower similarity to a target may produce better zero-shot transfer than uniform sampling.
  • If visual and RGB inputs were added to the benchmark, cross-embodiment generalization may interact with perception in ways that change the difficulty ordering, since the current setup isolates embodiment from vision by design.
  • A learned morphology representation inferred from the link sequence might replace the hand-crafted 48-dimensional vector, and whether such a learned encoding preserves the cosine-similarity ordering is a testable question for future work.
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. The paper introduces AnyBody, a simulated benchmark for cross-embodiment manipulation with reach and push tasks across a diverse set of robot morphologies, organized into interpolation, composition, and extrapolation regimes. The authors evaluate single-embodiment (SE) and multi-embodiment (ME) PPO agents with MLP and Transformer policies, reporting multi-task and zero-shot scores. They report that ME agents can match SE agents in-distribution but perform poorly on zero-shot generalization to novel morphologies, especially in extrapolation and composition, and identify design choices such as symlog returns and discrete actions that improve reach performance.

Significance. If validated, AnyBody would fill a real gap: existing cross-embodiment manipulation evaluations use only a handful of robots, while this benchmark provides a larger morphological space, clearly specified tasks, reward functions, and training configurations, plus an open-source IsaacSim extension. The paper's empirical finding that zero-shot generalization to novel morphologies remains unsolved is useful and not circular, since it is an experimental result rather than a consequence of fitted parameters. The ablation on symlog and discrete actions is also informative. However, the significance is currently limited by lack of validation of the difficulty axes, lack of statistical error bars, and a representation-dependent evaluation that the authors themselves acknowledge.

major comments (4)
  1. [Section 3.1, Figure 3, Table 3] The benchmark's central organizing claim is that interpolation, composition, and extrapolation form a difficulty gradient, but this is supported only by cosine similarity of the hand-crafted 48-dimensional link vectors. The reported cosine ranges overlap across regimes (interpolation: 0.80/0.76; composition: 0.47/0.49; extrapolation: 0.45/0.42), and the paper does not show that this metric predicts policy-transfer success. Table 3 also shows large differences in average link count and movable joints across regimes (e.g., Primitives 6.6 links vs. Robot Arms 14.7), so the observed zero-shot failures may reflect control difficulty or task complexity rather than morphological novelty. Please provide per-morphology validation of the cosine metric against actual zero-shot performance, or restrict the difficulty-ordering claims to what the evidence supports.
  2. [Section 5.1 and Section 5.2] All RL results appear to be single-run point estimates, with the reported score being the best score during up to 1M training steps. Under these conditions, differences such as those in Figures 6 and 8 cannot be distinguished from training noise, and statements such as "outperform" or "fail" are not statistically supported. Please run at least 3-5 seeds per condition, report mean and standard deviation, and use a fixed evaluation checkpoint (for example, the final checkpoint or the best checkpoint on a validation set) rather than the best training-time score.
  3. [Appendix 5, item 2] The paper concedes that the link-geometry approximation for real robots "might limit the model's performance." Since every evaluated agent consumes this approximate representation, the zero-shot generalization failures in the Robot Arms setting may be artifacts of the representation rather than evidence about morphology-aware learning methods. Please evaluate at least one agent with a more exact geometry encoding, or state explicitly in the main text which conclusions are representation-specific and which are expected to transfer to other morphology encodings.
  4. [Section 3.1, EE-Task; Section 5.2] In the EE-Task composition condition, the training set combines push on one morphology with reach on another, and the test is push on a new morphology. This setup conflates multi-task interference (push versus reach) with compositional generalization over morphology, as the authors partly acknowledge when they mention "inter-task complexity." The resulting zero-shot score cannot be attributed solely to failure to compose known body components. Please provide an ablation in which the same push task is trained on both component morphologies and tested on the composed morphology, or interpret EE-Task as a joint task-plus-embodiment generalization result rather than a pure morphology-composition result.
minor comments (5)
  1. [Section 5.2, Q1] The sentence claiming that "the Transformer-based ME agent consistently outperforms SE agents by a large margin" is not directly supported by Figure 6, where the plain Transformer bar is lower than the SE bar in some configurations; please clarify whether this sentence refers to the full Tf+Symlog+Discrete variant and label the figure bars accordingly.
  2. [Section 3.1] The 48-dimensional link-sequence vector is used in Figure 3 but is only fully defined in Table 4 of the appendix; consider summarizing its components in the main text so that the difficulty ordering is self-contained.
  3. [Section 5.1] The reach score is reported relative to a random agent, and several reported zero-shot scores are negative; please explain how to interpret a policy that is worse than random under this metric and whether this outcome reflects reward shaping, suboptimal local behavior, or the relative score's construction.
  4. [Figure 8 and Table 2] The abbreviations Sl, Dis, and TRW appear in the ablation table and figures without a caption-level definition; please add a legend or define them in the caption for readability.
  5. [Section 1 and Table 1] The text states that the benchmark includes 18 robot variations, but Table 1 lists six benchmark families; please reconcile this count explicitly, for example by listing the individual robot names and their assignment to train and test sets.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the benchmark results are empirical measurements, not derivations from fitted parameters or self-citations.

full rationale

The paper's central claims are empirical evaluations on a newly constructed benchmark, not derivations from fitted quantities. The three generalization regimes are defined by explicit structural criteria (same category, recombined components, different link structure), and the cosine-similarity figure in Sec. 3.1 is presented as a descriptive consistency check of those splits rather than as a predictor of the reported ZS/MT scores. Table 2 ablations and Figures 6–12 are direct measurements of RL training; no parameter is fitted to a subset and then 'predicted' on a closely related quantity. The only self-citation, FetchBench [30], appears in a list of simulation benchmarks in Related Work and is not load-bearing. The Appendix 5 admission that primitive-shape link approximations 'might limit the model's performance' is a validity caveat, not a circular dependency. Any concern that cosine similarity may not track policy transfer is a correctness/confounding question (e.g., Table 3 differences in link counts), not a circularity of the paper's argument.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

No new physical entities are postulated. The benchmark robots are simulated assets, not theoretical constructs. Free parameters are hand-set reward weights and thresholds that shape the tasks; they are part of the benchmark design, not fitted to force a result.

free parameters (5)
  • Reach reward weights (w1, w2, w3, w4) = Specified in Table 5 of Appendix; exact numeric values not in main text
    Hand-chosen weights for joint-limit penalty, joint-acc penalty, EE-goal distance, and vicinity reward; they shape agent behavior and task difficulty.
  • Push reward weights (w1, w2, w3) = Specified in Table 6 of Appendix; exact numeric values not in main text
    Hand-chosen weights for joint-limit penalty, object distance, and termination reward.
  • Success/vicinity distance thresholds = Not stated numerically in main text; referred to as 'threshold' in Tables 5 and 6
    Define when the task counts as solved; directly control reported success rates.
  • Discrete action bin spacing = Exponentially spaced bins around zero
    Chosen to simplify learning zero-output actions; affects control precision in the reach task.
  • Push curriculum goal threshold schedule = Gradually increasing y-value threshold
    Curriculum design influences training dynamics and final push success rates.
assumptions (4)
  • standard math Environments are infinite-horizon MDPs with shared horizon H and discount γ=0.99.
    Section 3.2; standard RL formalism assumed for all robot morphologies.
  • domain assumption A primitive-shape approximation (best-fit box, sphere, cylinder via Trimesh) adequately represents link geometry for policy learning.
    Appendix 2.1; used to construct the 48-dimensional link state; if approximation is poor, the morphologies are not faithfully encoded.
  • ad hoc to paper Cosine similarity of the 48-dimensional link sequence vectors ordering (interpolation < composition < extrapolation) reflects true generalization difficulty.
    Section 3.1; the benchmark's three axes and interpretations of results rest on this assumption.
  • domain assumption Simulation in IsaacSim is a faithful proxy for real-world manipulation physics.
    All experiments are simulated; real-world transfer is not validated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AnyBody: A Benchmark Suite for Cross-Embodiment Manipulation." pith.science (2026). https://pith.science/paper/ICGD2KMQ

@misc{pith2026250514986,
  author       = {Pith},
  title        = {Pith review of: AnyBody: A Benchmark Suite for Cross-Embodiment Manipulation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ICGD2KMQ}},
  note         = {Machine review of arXiv:2505.14986}
}
read the original abstract

Generalizing control policies to novel embodiments remains a fundamental challenge in enabling scalable and transferable learning in robotics. While prior works have explored this in locomotion, a systematic study in the context of manipulation tasks remains limited, partly due to the lack of standardized benchmarks. In this paper, we introduce a benchmark for learning cross-embodiment manipulation, focusing on two foundational tasks-reach and push-across a diverse range of morphologies. The benchmark is designed to test generalization along three axes: interpolation (testing performance within a robot category that shares the same link structure), extrapolation (testing on a robot with a different link structure), and composition (testing on combinations of link structures). On the benchmark, we evaluate the ability of different RL policies to learn from multiple morphologies and to generalize to novel ones. Our study aims to answer whether morphology-aware training can outperform single-embodiment baselines, whether zero-shot generalization to unseen morphologies is feasible, and how consistently these patterns hold across different generalization regimes. The results highlight the current limitations of multi-embodiment learning and provide insights into how architectural and training design choices influence policy generalization.

Figures

Figures reproduced from arXiv: 2505.14986 by the authors.

Figure 1
Figure 1. We introduce AnyBody, a benchmark suite for evaluating policy generalization across diverse robot morphologies. It consists of 18 robot variations: 8 procedurally generated robot categories and 10 based on real-world robots. The benchmark tasks comprise two manipulation tasks —reach and push— two scene variations (with and without obstacles), and two input types—state-based and point cloud-based. arms or multi-finge… view at source ↗
Figure 2
Figure 2. Three categories of benchmark tasks. We aim to test the zero-shot generaliza￾tion ability of a multi-embodiment policy π(s) on unseen morphologies. Framed as a multi-task learning problem, the bench￾mark evaluates a method’s ability to learn from multi￾embodiment data and generalize to novel morpholo￾gies. It tests generalization along three axes: interpola￾tion, extrapolation, and composition (see [PITH_FULL_IMAGE… view at source ↗
Figure 3
Figure 3. Cosine similarity of test morphologies with those in the train set. composition, some of the train robots have high similarity with the test robot, and for extrapolation, we observe that the test robot is farther from the train robots. 3.2 Modeling Following prior works [3, 8, 11], we model the different robotic environments as sets of infinite horizon, discounted Markov Decision Processes (MDPs) M = {M1, M2, · · · … view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Robot morphology is represented by a sequence of links. We approximate the link geometries by the shape param￾eters of a fitted primitive. Action Space. We use joint-space control as the action space, where the policy outputs joint position changes ∆q. Control in￾terfa…
Figure 5
Figure 5. Figure 5: Large￾scale multi￾embodiment training in Isaac-Sim. Morphology-aware training. We focus on two agent categories: single￾embodiment (SE) and multi-embodiment (ME). SE agents are trained sepa￾rately for each morphology using a dedicated actor-critic network. While they c…
Figure 6
Figure 6. Figure 6: MT score for (a) reach, and (b) push. The agents are: Single-Embodiment MLP Transformer [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Panda-push RL train￾ing with curriculum. We believe that ME training provides a regularizing effect, leading to more stable learning and reducing the chance of convergence to poor local minima. This can be seen in [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: ZS score for (a) reach, and (b) push. The agents are: Single-Embodiment MLP Transformer 7 [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]
Figure 9
Figure 9. Figure 9: Ablation: Train curves for different variations. A discrete action space leads to faster convergence. Discussion. The experiments highlight a key challenge: while Transformer agents match single-embodiment baselines in multi-task performance, they struggle with zero-sh…
Figure 10
Figure 10. Figure 10: Zero-shot and Finetuned model performance on test morphology for [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]
Figure 11
Figure 11. Figure 11: Zero-shot and Finetuned model performance on test morphology for [PITH_FULL_IMAGE:figures/full_fig_p016_11.png]
Figure 12
Figure 12. Figure 12: Performance (y-axis) on reach task with the number of fine-tuning steps (x-axis). Note that the performance of agents saturates in the Arm3 environments. These environments are designed to be simple, both to test models on simpler morphologies and to allow a faster tr…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 15 canonical work pages

  1. [1]

    A. O. et. al. Open x-embodiment: Robotic learning datasets and rt-x models : Open x- embodiment collaboration0.2024 IEEE International Conference on Robotics and Automation (ICRA), pages 6892–6903, 2024. URLhttps://arxiv.org/abs/2310.08864

    arXiv 2024

  2. [2]

    A. K. et. al. Droid: A large-scale in-the-wild robot manipulation dataset.ArXiv, abs/2403.12945, 2024

    arXiv 2024

  3. [3]

    Gupta, L

    A. Gupta, L. J. Fan, S. Ganguli, and L. Fei-Fei. Metamorph: Learning universal controllers with transformers.ArXiv, abs/2203.11931, 2022

    arXiv 2022

  4. [4]

    Patel and S

    A. Patel and S. Song. GET-Zero: Graph embodiment transformer for zero-shot embodiment generalization.2025 IEEE International Conference on Robotics and Automation (ICRA),

    work page 2025

  5. [5]

    J. Yang, C. Glossop, A. Bhorkar, D. Shah, Q. Vuong, C. Finn, D. Sadigh, and S. Levine. Pushing the limits of cross-embodiment learning for manipulation and navigation.ArXiv, abs/2402.19432, 2024

    arXiv 2024

  6. [8]

    T. Chen, A. Murali, and A. K. Gupta. Hardware conditioned policies for multi-robot transfer learning. InNeural Information Processing Systems, 2018

    work page 2018

  7. [9]

    E. S. Hu, K.-Y . Huang, O. Rybkin, and D. Jayaraman. Know thyself: Transferable visuomotor control through robot-awareness.ArXiv, abs/2107.09047, 2021

    arXiv 2021

  8. [10]

    Sferrazza, D.-M

    C. Sferrazza, D.-M. Huang, F. Liu, J. Lee, and P. Abbeel. Body transformer: Leveraging robot embodiment for policy learning.arXiv preprint arXiv:2408.06316, 2024. 9

    arXiv 2024

Show all 42 references
  1. [11]

    Kurin, M

    V . Kurin, M. Igl, T. Rockt¨aschel, W. Boehmer, and S. Whiteson. My body is a cage: the role of morphology in graph-based incompatible control. InInternational Conference on Learning Representations, 2021. URLhttps://openreview.net/forum?id=N3zUDGN5lO

    2021

  2. [12]

    T. Wang, R. Liao, J. Ba, and S. Fidler. Nervenet: Learning structured policy with graph neural networks. InInternational Conference on Learning Representations, 2018

    2018

  3. [13]

    J. Yang, D. Sadigh, and C. Finn. Polybot: Training one policy across robots while embracing variability. InConference on Robot Learning, 2023

    2023

  4. [14]

    T. Yu, D. Quillen, Z. He, R. Julian, A. Narayan, H. Shively, A. Bellathur, K. Hausman, C. Finn, and S. Levine. Meta-world: A benchmark and evaluation for multi-task and meta reinforce- ment learning, 2021. URLhttps://arxiv.org/abs/1910.10897

    2021 arXiv

  5. [15]

    Nvidia isaac sim, 2024

    NVIDIA Corporation. Nvidia isaac sim, 2024. Version 4.0. Available athttps:// developer.nvidia.com/isaac/sim

    2024

  6. [16]

    Mittal, C

    M. Mittal, C. Yu, Q. Yu, J. Liu, N. Rudin, D. Hoeller, J. L. Yuan, R. Singh, Y . Guo, H. Mazhar, A. Mandlekar, B. Babich, G. State, M. Hutter, and A. Garg. Orbit: A unified simulation framework for interactive robot learning environments.IEEE Robotics and Automation Letters, 8...

    2023

  7. [17]

    Makoviychuk et al

    V . Makoviychuk et al. Isaac lab: A unified framework for robot learning.https://github. com/isaac-sim/IsaacLab, 2022

    2022

  8. [18]

    Bjorck, F

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

    2025 arXiv

  9. [19]

    Black, N

    K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Haus- man, B. Ichter, S. Jakubczak, T. Jones, L. Ke, S. Levine, A. Li-Bell, M. Mothukuri, S. Nair, K. Pertsch, L. X. Shi, J. Tanner, Q. Vuong, A. Walling, H. Wang, and U. Zhilinsky.π 0: A vision...

    2024 arXiv

  10. [20]

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

    2024 arXiv

  11. [21]

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

    2024 arXiv

  12. [22]

    Bousmalis, G

    K. Bousmalis, G. Vezzani, D. Rao, C. Devin, A. X. Lee, M. Bauz ´a, T. Davchev, Y . Zhou, A. Gupta, A. Raju, et al. Robocat: A self-improving generalist agent for robotic manipulation. arXiv preprint arXiv:2306.11706, 2023

    2023 arXiv

  13. [23]

    S. Reed, K. Zolna, E. Parisotto, S. G. Colmenarejo, A. Novikov, G. Barth-Maron, M. Gim´enez, Y . Sulsky, J. Kay, J. T. Springenberg, T. Eccles, J. Bruce, A. Razavi, A. D. Edwards, N. M. O. Heess, Y . Chen, R. Hadsell, O. Vinyals, M. Bordbar, and N. de Freitas. A generalist age...

    2022 arXiv

  14. [24]

    J. Gu, F. Xiang, X. Li, Z. Ling, X. Liu, T. Mu, Y . Tang, S. Tao, X. Wei, Y . Yao, et al. Maniskill2: A unified benchmark for generalizable manipulation skills.arXiv preprint arXiv:2302.04659, 2023

    2023 arXiv

  15. [25]

    Ehsani, W

    K. Ehsani, W. Han, A. Herrasti, E. VanderBilt, L. Weihs, E. Kolve, A. Kembhavi, and R. Mot- taghi. Manipulathor: A framework for visual object manipulation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4497–4506, 2021. 10

    2021

  16. [26]

    James, Z

    S. James, Z. Ma, D. R. Arrojo, and A. J. Davison. Rlbench: The robot learning benchmark & learning environment.IEEE Robotics and Automation Letters, 5(2):3019–3026, 2020

    2020

  17. [27]

    Y . Zhu, J. Wong, A. Mandlekar, R. Mart ´ın-Mart´ın, A. Joshi, S. Nasiriany, and Y . Zhu. ro- bosuite: A modular simulation framework and benchmark for robot learning.arXiv preprint arXiv:2009.12293, 2020

    2009 arXiv

  18. [28]

    T. Yu, D. Quillen, Z. He, R. Julian, K. Hausman, C. Finn, and S. Levine. Meta-world: A benchmark and evaluation for multi-task and meta reinforcement learning. InConference on robot learning, pages 1094–1100. PMLR, 2020

    2020

  19. [29]

    M. Heo, Y . Lee, D. Lee, and J. J. Lim. Furniturebench: Reproducible real-world benchmark for long-horizon complex manipulation. InRobotics: Science and Systems, 2023

    2023

  20. [30]

    B. Han, M. Parakh, D. Geng, J. A. Defay, G. Luyang, and J. Deng. Fetchbench: A simulation benchmark for robot fetching.arXiv preprint arXiv:2406.11793, 2024

    2024 arXiv

  21. [31]

    C. Li, R. Zhang, J. Wong, C. Gokmen, S. Srivastava, R. Mart ´ın-Mart´ın, C. Wang, G. Levine, M. Lingelbach, J. Sun, et al. Behavior-1k: A benchmark for embodied ai with 1,000 everyday activities and realistic simulation. InConference on Robot Learning, pages 80–93. PMLR, 2023

    2023

  22. [32]

    Srivastava, C

    S. Srivastava, C. Li, M. Lingelbach, R. Mart ´ın-Mart´ın, F. Xia, K. E. Vainio, Z. Lian, C. Gok- men, S. Buch, K. Liu, et al. Behavior: Benchmark for everyday household activities in virtual, interactive, and ecological environments. InConference on robot learning, pages 477–4...

    2022

  23. [33]

    A. Szot, A. Clegg, E. Undersander, E. Wijmans, Y . Zhao, J. Turner, N. Maestre, M. Mukadam, D. S. Chaplot, O. Maksymets, et al. Habitat 2.0: Training home assistants to rearrange their habitat.Advances in neural information processing systems, 34:251–266, 2021

    2021

  24. [34]

    Sferrazza, D.-M

    C. Sferrazza, D.-M. Huang, X. Lin, Y . Lee, and P. Abbeel. Humanoidbench: Simulated hu- manoid benchmark for whole-body locomotion and manipulation, 2024

    2024

  25. [35]

    Gupta, S

    A. Gupta, S. Savarese, S. Ganguli, and L. Fei-Fei. Embodied intelligence via learning and evolution.Nature Communications, 12(1), Oct. 2021. ISSN 2041-1723. doi:10.1038/ s41467-021-25874-z. URLhttp://dx.doi.org/10.1038/s41467-021-25874-z

    2021 doi

  26. [36]

    S. Hong, D. Yoon, and K.-E. Kim. Structure-aware transformer policy for inhomogeneous multi-task reinforcement learning. InInternational Conference on Learning Representations, 2022

    2022

  27. [37]

    Zhang, T

    Q. Zhang, T. Xiao, A. A. Efros, L. Pinto, and X. Wang. Learning cross-domain correspondence for control with dynamics cycle-consistency.arXiv preprint arXiv:2012.09811, 2020

    2012 arXiv

  28. [38]

    Schulman, F

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov. Proximal policy optimization algorithms.ArXiv, abs/1707.06347, 2017

    2017 arXiv

  29. [39]

    Vaswani, N

    A. Vaswani, N. M. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin. Attention is all you need. InNeural Information Processing Systems, 2017

    2017

  30. [40]

    Hafner, J

    D. Hafner, J. Pasukonis, J. Ba, and T. Lillicrap. Mastering diverse domains through world models, 2024. URLhttps://arxiv.org/abs/2301.04104

    2024 arXiv

  31. [41]

    Brohan, N

    A. Brohan, N. Brown, J. Carbajal, Y . Chebotar, J. Dabis, C. Finn, K. Gopalakrishnan, K. Haus- man, A. Herzog, J. Hsu, et al. Rt-1: Robotics transformer for real-world control at scale.arXiv preprint arXiv:2212.06817, 2022. 11

    2022 arXiv

  32. [42]

    Z. Yuan, T. Wei, S. Cheng, G. Zhang, Y . Chen, and H. Xu. Learning to manipulate anywhere: A visual generalizable framework for reinforcement learning. In8th Annual Conference on Robot Learning, 2024. URLhttps://openreview.net/forum?id=jart4nhCQr

    2024

  33. [43]

    Serrano-Mu ˜noz, D

    A. Serrano-Mu ˜noz, D. Chrysostomou, S. Bøgh, and N. Arana-Arexolaleiba. skrl: Modular and flexible library for reinforcement learning.Journal of Machine Learning Research, 24(254): 1–9, 2023. URLhttp://jmlr.org/papers/v24/23-0112.html. 12 Appendix 1 Benchmark Details 1.1 Benc...

    2023

  34. [2024]

    URLhttps://arxiv.org/abs/2407.15002

    arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.