Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Adding a plug-in equivariance loss that makes bimanual policies treat left-right mirrored observations consistently improves success and robustness across vision, action spaces, and real-robot shifts—without changing the network architectur

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 18:30 UTC pith:K5WPRKS7

load-bearing objection A clean output-level equivariance regularizer with real average gains — but the joint-space symmetry map S(q)=Dq is asserted, never validated, and 'consistently improves' overreaches Table II. the 3 major comments →

arxiv 2603.08541 v3 pith:K5WPRKS7 submitted 2026-03-09 cs.RO

EquiBim: Learning Symmetry-Equivariant Policy for Bimanual Manipulation

Zhiyuan Zhang , Aditya Mohan , Seungho Han , Wan Shou , Dongyi Wang , Yu She This is my paper
classification cs.RO
keywords symmetry-equivariant policybimanual manipulationimitation learningbilateral symmetryequivariance regularizationbehavior cloningdistribution shiftdual-arm robot
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.

EquiBim claims that many failures and inconsistencies in dual-arm imitation learning stem from ignoring the bilateral symmetry of the robot and task. The paper's central proposal is a single regularization term, L_sym = ||π(S(O)) − S(π(O))||², added to any existing behavior-cloning objective. This term requires the policy's predictions on a left-right reflected observation to equal the reflected predictions on the original observation, effectively teaching the policy to be symmetric without architectural modifications. Across image- and point-cloud-based policies with end-effector and joint-space action representations, the approach raises average success rates on an eight-task benchmark by up to 9.5 points and improves robustness under distribution shifts in real-robot trials. A sympathetic reader would take the paper to establish that explicit output-level symmetry enforcement is a simple, model-agnostic inductive bias worth adding to bimanual pipelines.

Core claim

The central claim is that enforcing bilateral equivariance between observations and actions during training, through the action-level consistency loss L_sym = ||π(S(O)) − S(π(O))||², makes imitation-learned bimanual policies more accurate and more robust to distribution shifts. EquiBim defines a symmetry transformation S that swaps the robot's left and right sides while reflecting visual inputs: a horizontal flip for images, a lateral-axis reflection for point clouds, and for joint-space actions a fixed diagonal sign-flip matrix derived from the robot's URDF. The regularization is applied purely at the output level and leaves the policy architecture, optimizer, and pipeline untouched, so any

What carries the argument

The equivariance loss L_sym = ||π(S(O)) − S(π(O))||² is the mechanism that carries the argument. It is an output-level consistency constraint: given an observation O and its symmetrically transformed counterpart S(O), the policy is penalized for producing predictions that do not transform the same way under S. The symmetry transformation S itself is the second load-bearing piece; per modality it is instantiated as a horizontal image flip, a point-cloud reflection through the camera frame, or a joint-space sign-flip matrix D. The work that S does is to turn the abstract prior 'left-right exchange is a valid task symmetry' into a concrete, differentiable training signal compatible with any pol

Load-bearing premise

The load-bearing premise is that exchanging the two arms is exactly equivalent to a fixed coordinate reflection: a horizontal image flip, a point-cloud reflection, and, in joint space, a constant diagonal sign-flip matrix derived once from the robot model; for real articulated arms the true mapping between left-arm and right-arm joint configurations generally involves offsets and nonlinear kinematics, so this fixed map can be systematically wrong for the joint-space and certa

What would settle it

Take the robot's URDF, pick a random joint configuration q of the left arm, compute the true symmetric right-arm configuration by solving the kinematics of arm exchange, and compare it to Dq; if the residual is large, the equivariance constraint enforces an incorrect symmetry and the claimed gains would vanish or reverse, and the same could be tested by placing the camera off-center so a horizontal flip no longer coincides with the lateral reflection.

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

If this is right

  • Adding EquiBim to diffusion-policy and 3D-diffusion-policy baselines raises average success rates in all four observation-action settings on the simulation benchmark, with the largest gain (+9.5 points) on image-plus-joint policies where geometric cues are weakest.
  • On a real dual-arm platform, the regularization lifts banana-handover success from 3/10 to 6/10 under the training distribution and from 0/10 to 5/10 when the object's orientation is mirrored, indicating better symmetry transfer.
  • When demonstration quality is asymmetric across the two arms, the symmetry loss lets the stronger side's demonstrations regularize the weaker side, improving shifted-distribution hook-hanging success from 1/10 to 4/10.
  • The constraint can suppress useful asymmetry: two of the eight simulation tasks show performance drops because their optimal strategies are instance-level asymmetric (functional role asymmetry, grasp-order differences), so practitioners should check task-level symmetry before applying it.
  • Because the method is architecture-agnostic, any existing imitation learning pipeline for bimanual tasks can incorporate the loss by defining S for its own observation and action spaces.

Where Pith is reading between the lines

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

  • If the D-matrix derivation is exact, the same output-level equivariance idea could be applied to other discrete symmetries in manipulation, such as front-back reflections or rotational symmetries of the workspace, by swapping S accordingly; the paper does not test these.
  • The mechanism suggests a testable extension: instead of fixing S as a hard linear map, one could estimate the arm-exchange mapping online from data (a learned S), which might recover the losses on the two asymmetric tasks where a rigid sign-flip is systematically wrong.
  • The success on image-plus-joint suggests that equivariance regularization may be a cheap substitute for explicit 3D structure in visuomotor policies; a controlled study ablating point-cloud input against image-plus-EquiBim would test whether the loss closes the geometry gap.
  • The real-robot results cover only three tasks with 50 demonstrations each; the paper's positive headline claim would be more firmly established by showing it holds across a wider variety of bimanual tasks and demonstration counts.

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

3 major / 4 minor

Summary. The paper proposes EquiBim, a training-time regularization method for imitation learning policies in bimanual manipulation. The method defines a bilateral symmetry transformation S that acts consistently on observations and actions, and adds an equivariance loss L_sym = ||π(S(O)) − S(π(O))||² to the standard behavior-cloning objective. EquiBim is claimed to be model-agnostic and architecture-free, applicable to image- and point-cloud-based policies with end-effector or joint-space action outputs. The authors evaluate on eight RoboTwin simulation tasks across four observation/action configurations and on a real-world two-arm LeRobot platform with ACT, reporting average gains in simulation and improved robustness under distribution shifts in real-world trials. The paper's central claim is that EquiBim consistently improves performance and robustness without architectural changes.

Significance. If the central claim held, EquiBim would be a practically valuable contribution: it is a simple, plug-and-play training objective that could be added to any imitation learning pipeline for symmetric bimanual tasks, with no architectural modification. The breadth of the evaluation — four observation/action combinations in simulation and three real-world tasks with distribution shifts — is a genuine strength, as is the explicit attempt to define the symmetry transform across image, point cloud, and joint/end-effector spaces. However, the paper's evidence is weaker than its wording: the joint-space symmetry map in Eq. (3) is asserted rather than validated, per-task results in Table II include several regressions, and no error bars or significance tests are reported. The central idea is defensible and worth pursuing, but the current presentation overstates the consistency of the improvements.

major comments (3)
  1. [Sec. III-C, Eq. (3)] The joint-space symmetry map is defined as S(q) = Dq with a constant diagonal sign-flip matrix D, 'determined from the robot URDF.' This is a strong kinematic assumption. For articulated arms mounted symmetrically, exchanging the left and right arms generally maps joint angles through an affine map with joint offsets, q ↦ D(q − q0) + q0, or through a configuration-dependent transform, not simply a fixed sign flip. The paper never reports D, never derives it from the kinematic tree, and never validates it against the actual left–right mapping. This is load-bearing because all joint-space experiments (Image+Joint, Point Cloud+Joint, and the real-world ACT trials, which use joint-space actions) rely on Eq. (3). If D is wrong, L_sym enforces an incorrect equivariance, and the per-task drops in Table II (e.g., Handover Block −11 on Image+Joint, Pick Dual Bottles −13 on Image+EE) could be caus
  2. [Sec. IV-B, Tables I and II] The central claim is that EquiBim 'consistently improves performance and robustness.' However, Table II shows several negative per-task deltas (e.g., Handover Block, Pick Dual Bottles, and multiple Point Cloud entries), some as large as −15. These regressions are acknowledged in the text but are not reconciled with the word 'consistently.' Moreover, no error bars, standard deviations, number of seeds, or statistical tests are reported anywhere; Table I reports only average success rates over what appears to be a single seed per setting, and Table III reports 10 trials per condition with no confidence intervals. With 10 Bernoulli trials, 3/10 vs 6/10 is not strong evidence of a 30% improvement. The authors should report multiple seeds with means and variances, provide per-task error bars, and either temper the 'consistently improves' claim or support it with a paired statistical test acro
  3. [Sec. III-B and Sec. IV-A] The loss L_sym in Eq. (1) includes an implicit scalar weighting relative to the imitation loss, but the weight value is never reported and no sensitivity analysis or ablation is provided. Since all reported improvements depend on this hyperparameter, the paper is not reproducible without it. Please report the exact training objective (including the weight) and provide an ablation over at least a few weight values for one or two settings. This is a local fix, but it is necessary for the method to be assessed fairly.
minor comments (4)
  1. [Sec. IV-B, Table II] The table header 'Place' corresponds to 'Place Empty Cup' in Fig. 3, but the row/column alignment is easy to misread. Consider using full task names or a clear legend.
  2. [Sec. IV-C, Fig. 6] The visualization says 'only a subset of representative rollouts is displayed,' but the selection criterion is not stated. If rollouts are hand-picked, the figure should be replaced with all rollouts or with summary statistics, to avoid cherry-picking concerns.
  3. [Sec. IV-C, Table III] The real-world evaluation uses only 10 trials per condition and a single camera/robot setup. Please state how trials were randomized and whether any trials were discarded; also provide the raw per-trial outcomes or at least binomial confidence intervals.
  4. [General] The project website URL is not hyperlinked in the PDF and appears broken in the text ('https:zhangzhiyuanzhang.github.io/equibim-website/'). Please correct the URL and consider including a code release link or an appendix with training details.

Circularity Check

0 steps flagged

No significant circularity: the equivariance loss directly defines the enforced property, and the reported gains are empirical measurements rather than derived predictions.

full rationale

The paper's core objective, L_sym = ||π(S(O)) − S(π(O))||^2 (Eq. 1), is defined as the squared violation of the equivariance condition itself; thus the statement that EquiBim enforces bilateral equivariance is true by construction, but this is a design definition, not a derived prediction. No parameter is fitted to the evaluation data and later renamed a prediction: the symmetry transforms for images, point clouds, and end-effector poses are geometric reflections, and the joint-space transform S(q)=Dq (Eq. 3) is asserted from the URDF rather than fit to task outcomes. The improvements in Tables I–III are empirical results, not algebraic consequences of the loss. The main caveats—the unvalidated constant diagonal D and the selection of tasks and shifts that are symmetric by construction—are correctness and experimental-design concerns, not circularity: if D is wrong, the enforced equivariance is false, but the derivation does not reduce to its inputs. Self-citations [11] and [16] appear only in related-work context and are not load-bearing for the framework's validity. Therefore, no circular step is exhibited, and the appropriate finding is no significant circularity.

Axiom & Free-Parameter Ledger

1 free parameters · 4 axioms · 0 invented entities

The central claim depends on a strong task-symmetry prior plus a linear joint-space model of symmetry. No fitted constants appear in the derivation, but the loss weight is unreported and the joint sign matrix is assumed exact. The symmetry transformation S is a mathematical construction, not an empirical entity.

free parameters (1)
  • symmetry loss weight = not reported
    The combined training loss weight for L_sym is not stated anywhere; if tuned per task or per run, it could inflate measured gains.
axioms (4)
  • domain assumption Expert demonstrations for the selected tasks are exactly equivariant under the left-right symmetry S.
    Section III-A asserts a 'task-level prior' that exchanging arms yields equally valid expert behavior; if a task has instance-level asymmetries, enforcing this prior hurts performance, as the paper itself reports in Table II.
  • domain assumption A centered head camera makes the image horizontal axis coincide with the robot's lateral direction, so S is a horizontal flip.
    Section III-C: all image and point-cloud transforms assume this alignment; small camera offsets would break the supposed symmetry.
  • domain assumption Joint-space symmetry is exactly represented by a fixed constant diagonal sign matrix D.
    Eq. (3): S(q)=Dq; for articulated two-arm systems this is generally configuration-dependent, and the paper does not validate D.
  • domain assumption Standard behavior-cloning assumptions hold: demonstrations are representative and the policy class can represent the expert.
    The whole evaluation is behavior cloning from demonstrations; failure of this assumption would affect both baseline and EquiBim, but it is still load-bearing for the reported success rates.

pith-pipeline@v1.3.0-alltime-deepseek · 9264 in / 9496 out tokens · 90967 ms · 2026-08-02T18:30:47.180933+00:00 · methodology

0 comments
read the original abstract

Robotic imitation learning has achieved impressive success in learning complex manipulation behaviors from demonstrations. However, many existing robot learning methods do not explicitly account for the physical symmetries of robotic systems, often resulting in asymmetric or inconsistent behaviors under symmetric observations. This limitation is particularly pronounced in dual-arm manipulation, where bilateral symmetry is inherent to both the robot morphology and the structure of many tasks. In this paper, we introduce EquiBim, a symmetry-equivariant policy learning framework for bimanual manipulation that enforces bilateral equivariance between observations and actions during training. Our approach formulates physical symmetry as a group action on both observation and action spaces, and imposes an equivariance constraint on policy predictions under symmetric transformations. The framework is model-agnostic and can be seamlessly integrated into a wide range of imitation learning pipelines with diverse observation modalities and action representations, including point cloud-based and image-based policies, as well as both end-effector-space and joint-space parameterizations. We evaluate EquiBim on RoboTwin, a dual-arm robotic platform with symmetric kinematics, and evaluate it across diverse observation and action configurations in simulation. We further validate the approach on a real-world dual-arm system. Across both simulation and physical experiments, our method consistently improves performance and robustness under distribution shifts. These results suggest that explicitly enforcing physical symmetry provides a simple yet effective inductive bias for bimanual robot learning.

Figures

Figures reproduced from arXiv: 2603.08541 by Aditya Mohan, Dongyi Wang, Seungho Han, Wan Shou, Yu She, Zhiyuan Zhang.

Figure 1
Figure 1. Figure 1: Symmetry-equivariant policy learning for bimanual manipu￾lation. A symmetry transformation S defines a left–right exchange of the scene in the image coordinate frame, generating a symmetrically equivalent task instance. Given an observation O and its transformed counterpart S(O), the shared policy π is trained to produce equivariant predictions under the same transformation, i.e., π(S(O)) ≈ S(π(O)). The sy… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of EquiBim. Given an observation [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visualization of eight simulated bimanual manipulation tasks used for evaluation. Subfigures (a)-(g) are adapted from RoboTwin [19]. For each [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overview of the real-world bimanual manipulation setup. Two [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Real-world task executions on the bimanual LeRobot platform. The figure presents three tasks, namely Object Handover (Banana), Hook Hanging [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Generalization under distribution shift on the real-world bimanual [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] 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 1 Pith paper

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

  1. Morphologically Equivariant Flow Matching for Bimanual Mobile Manipulation

    cs.RO 2026-05 conditional novelty 7.0

    A morphologically equivariant flow matching policy for bimanual robots enforces reflective symmetry to improve sample efficiency and enable zero-shot generalization to mirrored task configurations.

Reference graph

Works this paper leans on

31 extracted references · 10 linked inside Pith · cited by 1 Pith paper

  1. [1]

    An algorithmic perspective on imitation learning.F oundations and Trends® in Robotics, 7(1-2):1–179, 2018

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

    2018

  2. [2]

    Diffusion policy: Visuomotor policy learning via action diffusion

    Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. InProceedings of Robotics: Science and Systems, 2023

    2023

  3. [3]

    Learning fine-grained bimanual manipulation with low-cost hardware

    Tony Z Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manipulation with low-cost hardware. InProceedings of Robotics: Science and Systems, 2023

    2023

  4. [4]

    Aloha unleashed: A simple recipe for robot dexterity.arXiv preprint arXiv:2410.13126, 2024

    Tony Z Zhao, Jonathan Tompson, Danny Driess, Pete Florence, Kam- yar Ghasemipour, Chelsea Finn, and Ayzaan Wahid. Aloha unleashed: A simple recipe for robot dexterity.arXiv preprint arXiv:2410.13126, 2024

    Pith/arXiv arXiv 2024

  5. [5]

    Dexterous manipulation through imitation learning: A survey

    Shan An, Ziyu Meng, Chao Tang, Yuning Zhou, Tengyu Liu, Fangqiang Ding, Shufang Zhang, Yao Mu, Ran Song, Wei Zhang, et al. Dexterous manipulation through imitation learning: A survey. arXiv preprint arXiv:2504.03515, 2025

    arXiv 2025

  6. [6]

    Tacdiffusion: Force-domain diffusion policy for precise tactile ma- nipulation

    Yansong Wu, Zongxie Chen, Fan Wu, Lingyun Chen, Liding Zhang, Zhenshan Bing, Abdalla Swikir, Sami Haddadin, and Alois Knoll. Tacdiffusion: Force-domain diffusion policy for precise tactile ma- nipulation. In2025 IEEE International Conference on Robotics and Automation (ICRA), pages 11831–11837. IEEE, 2025

    2025

  7. [7]

    Vtla: Vision-tactile-language-action model with preference learning for insertion manipulation.arXiv preprint arXiv:2505.09577, 2025

    Chaofan Zhang, Peng Hao, Xiaoge Cao, Xiaoshuai Hao, Shaowei Cui, and Shuo Wang. Vtla: Vision-tactile-language-action model with preference learning for insertion manipulation.arXiv preprint arXiv:2505.09577, 2025

    Pith/arXiv arXiv 2025

  8. [8]

    Manifeel: Benchmarking and under- standing visuotactile manipulation policy learning.arXiv preprint arXiv:2505.18472, 2025

    Quan Khanh Luu, Pokuang Zhou, Zhengtong Xu, Zhiyuan Zhang, Qiang Qiu, and Yu She. Manifeel: Benchmarking and under- standing visuotactile manipulation policy learning.arXiv preprint arXiv:2505.18472, 2025

    arXiv 2025

  9. [9]

    Vitacformer: Learning cross-modal representation for visuo- tactile dexterous manipulation.arXiv preprint arXiv:2506.15953, 2025

    Liang Heng, Haoran Geng, Kaifeng Zhang, Pieter Abbeel, and Jitendra Malik. Vitacformer: Learning cross-modal representation for visuo- tactile dexterous manipulation.arXiv preprint arXiv:2506.15953, 2025

    Pith/arXiv arXiv 2025

  10. [10]

    Dexmimicgen: Automated data generation for bimanual dexterous manipulation via imitation learning

    Zhenyu Jiang, Yuqi Xie, Kevin Lin, Zhenjia Xu, Weikang Wan, Ajay Mandlekar, Linxi Jim Fan, and Yuke Zhu. Dexmimicgen: Automated data generation for bimanual dexterous manipulation via imitation learning. In2025 IEEE International Conference on Robotics and Automation (ICRA), pages 16923–16930. IEEE, 2025

    2025

  11. [11]

    Canonical policy: Learning canonical 3d representation for equivariant policy.arXiv preprint arXiv:2505.18474, 2025

    Zhiyuan Zhang, Zhengtong Xu, Jai Nanda Lakamsani, and Yu She. Canonical policy: Learning canonical 3d representation for equivariant policy.arXiv preprint arXiv:2505.18474, 2025

    arXiv 2025

  12. [12]

    A comparison of imitation learning algorithms for bimanual manipulation.IEEE Robotics and Automation Letters (RA-L), 2024

    Michael Drolet, Simon Stepputtis, Siva Kailas, Ajinkya Jain, Jan Peters, Stefan Schaal, and Heni Ben Amor. A comparison of imitation learning algorithms for bimanual manipulation.IEEE Robotics and Automation Letters (RA-L), 2024

    2024

  13. [13]

    Morphologically symmetric reinforce- ment learning for ambidextrous bimanual manipulation.arXiv preprint arXiv:2505.05287, 2025

    Zechu Li, Yufeng Jin, Daniel Ordonez Apraez, Claudio Semini, Puze Liu, and Georgia Chalvatzaki. Morphologically symmetric reinforce- ment learning for ambidextrous bimanual manipulation.arXiv preprint arXiv:2505.05287, 2025

    Pith/arXiv arXiv 2025

  14. [14]

    Morphological symmetries in robotics.The International Journal of Robotics Research, page 02783649241282422, 2025

    Daniel Ordonez Apraez, Giulio Turrisi, Vladimir Kostic, Mario Mar- tin, Antonio Agudo, Francesc Moreno-Noguer, Massimiliano Pontil, Claudio Semini, and Carlos Mastalli. Morphological symmetries in robotics.The International Journal of Robotics Research, page 02783649241282422, 2025

    2025

  15. [15]

    Equivariant diffusion policy

    Dian Wang, Stephen Hart, David Surovik, Tarik Kelestemur, Haojie Huang, Haibo Zhao, Mark Yeatman, Jiuguang Wang, Robin Walters, and Robert Platt. Equivariant diffusion policy. InProceedings of Conference on Robot Learning, 2024

    2024

  16. [16]

    Equiform: Noise-robust se (3)- equivariant policy learning from 3d point clouds.arXiv preprint arXiv:2601.17486, 2026

    Zhiyuan Zhang and Yu She. Equiform: Noise-robust se (3)- equivariant policy learning from 3d point clouds.arXiv preprint arXiv:2601.17486, 2026

    arXiv 2026

  17. [17]

    Roboagent: Generalization and efficiency in robot manipulation via semantic augmentations and action chunking.arXiv preprint arXiv:2309.01918, 2023

    Homanga Bharadhwaj, Jay Vakil, Mohit Sharma, Abhinav Gupta, Shubham Tulsiani, and Vikash Kumar. Roboagent: Generalization and efficiency in robot manipulation via semantic augmentations and action chunking.arXiv preprint arXiv:2309.01918, 2023

    Pith/arXiv arXiv 2023

  18. [18]

    Rovi-aug: Robot and viewpoint augmentation for cross-embodiment robot learning.arXiv preprint arXiv:2409.03403, 2024

    Lawrence Yunliang Chen, Chenfeng Xu, Karthik Dharmarajan, Muhammad Zubair Irshad, Richard Cheng, Kurt Keutzer, Masayoshi Tomizuka, Quan Vuong, and Ken Goldberg. Rovi-aug: Robot and viewpoint augmentation for cross-embodiment robot learning.arXiv preprint arXiv:2409.03403, 2024

    Pith/arXiv arXiv 2024

  19. [19]

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

    Tianxing Chen, Zanxin Chen, Baijun Chen, Zijian Cai, Yibin Liu, Qiwei Liang, Zixuan Li, Xianliang Lin, Yiheng Ge, Zhenyu Gu, et 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

    Pith/arXiv arXiv 2025

  20. [20]

    Dual arm manipulation—a survey.Robotics and Autonomous systems, 60(10):1340–1353, 2012

    Christian Smith, Yiannis Karayiannidis, Lazaros Nalpantidis, Xavi Gratal, Peng Qi, Dimos V Dimarogonas, and Danica Kragic. Dual arm manipulation—a survey.Robotics and Autonomous systems, 60(10):1340–1353, 2012

    2012

  21. [21]

    Bimanual dexterity for complex tasks.arXiv preprint arXiv:2411.13677, 2024

    Kenneth Shaw, Yulong Li, Jiahui Yang, Mohan Kumar Srirama, Ray Liu, Haoyu Xiong, Russell Mendonca, and Deepak Pathak. Bimanual dexterity for complex tasks.arXiv preprint arXiv:2411.13677, 2024

    Pith/arXiv arXiv 2024

  22. [22]

    Deep imitation learning for bimanual robotic manipulation.Advances in neural information processing systems, 33:2327–2337, 2020

    Fan Xie, Alexander Chowdhury, M De Paolis Kaluza, Linfeng Zhao, Lawson Wong, and Rose Yu. Deep imitation learning for bimanual robotic manipulation.Advances in neural information processing systems, 33:2327–2337, 2020

    2020

  23. [23]

    Anypos: Automated task-agnostic actions for bimanual manipulation.arXiv preprint arXiv:2507.12768, 2025

    Hengkai Tan, Yao Feng, Xinyi Mao, Shuhe Huang, Guodong Liu, Zhongkai Hao, Hang Su, and Jun Zhu. Anypos: Automated task-agnostic actions for bimanual manipulation.arXiv preprint arXiv:2507.12768, 2025

    Pith/arXiv arXiv 2025

  24. [24]

    A collaborative control method of dual-arm robots based on deep reinforcement learning.Applied Sciences, 11(4):1816, 2021

    Luyu Liu, Qianyuan Liu, Yong Song, Bao Pang, Xianfeng Yuan, and Qingyang Xu. A collaborative control method of dual-arm robots based on deep reinforcement learning.Applied Sciences, 11(4):1816, 2021

    2021

  25. [25]

    Equibot: SIM(3)-equivariant diffusion policy for generalizable and data efficient learning

    Jingyun Yang, Zi-ang Cao, Congyue Deng, Rika Antonova, Shuran Song, and Jeannette Bohg. Equibot: SIM(3)-equivariant diffusion policy for generalizable and data efficient learning. InProceedings of Conference on Robot Learning, 2024

    2024

  26. [26]

    Actionflow: Equivariant, accurate, and efficient policies with spatially symmetric flow matching.arXiv preprint arXiv:2409.04576, 2024

    Niklas Funk, Julen Urain, Joao Carvalho, Vignesh Prasad, Georgia Chalvatzaki, and Jan Peters. Actionflow: Equivariant, accurate, and efficient policies with spatially symmetric flow matching.arXiv preprint arXiv:2409.04576, 2024

    Pith/arXiv arXiv 2024

  27. [27]

    Highly accurate protein structure prediction with alphafold.nature, 596(7873):583–589, 2021

    John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin ˇZ´ıdek, Anna Potapenko, et al. Highly accurate protein structure prediction with alphafold.nature, 596(7873):583–589, 2021

    2021

  28. [28]

    Residual rotation correction using tactile equivariance

    Yizhe Zhu, Zhang Ye, Boce Hu, Haibo Zhao, Yu Qi, Dian Wang, and Robert Platt. Residual rotation correction using tactile equivariance. arXiv preprint arXiv:2511.07381, 2025

    arXiv 2025

  29. [29]

    E(2)-equivariant graph planning for navigation.IEEE Robotics and Automation Letters, 9(4):3371–3378, 2024

    Linfeng Zhao, Hongyu Li, Tas ¸kın Padır, Huaizu Jiang, and Lawson LS Wong. E(2)-equivariant graph planning for navigation.IEEE Robotics and Automation Letters, 9(4):3371–3378, 2024

    2024

  30. [30]

    3D Diffusion Policy: Generalizable visuomotor policy learning via simple 3d representations

    Yanjie Ze, Gu Zhang, Kangning Zhang, Chenyuan Hu, Muhan Wang, and Huazhe Xu. 3D Diffusion Policy: Generalizable visuomotor policy learning via simple 3d representations. InProceedings of Robotics: Science and Systems, 2024

    2024

  31. [31]

    Lerobot: An open-source library for end-to-end robot learning.arXiv preprint arXiv:2602.22818, 2026

    Remi Cadene, Simon Aliberts, Francesco Capuano, Michel Aractingi, Adil Zouitine, Pepijn Kooijmans, Jade Choghari, Martino Russi, Caroline Pascal, Steven Palma, et al. Lerobot: An open-source library for end-to-end robot learning.arXiv preprint arXiv:2602.22818, 2026

    arXiv 2026