Pith. sign in

REVIEW 3 major objections 6 minor 2 cited by

Benchmarking Generalizable Bimanual Manipulation: RoboTwin Dual-Arm Collaboration Challenge at CVPR 2025 MEIS Workshop

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A 64-team, three-stage competition benchmarked dual-arm manipulation across simulation and real robots and found that explicit 3D representations, data quality, and model capacity drive generalization.

desk verdict Useful competition record, but the real-world round's seen-configuration trials and released demos undercut the generalization claim; per-condition scores are needed. read the letter →

arxiv 2506.23351 v2 pith:UGL3W3XL submitted 2025-06-29 cs.RO cs.AIcs.LGcs.MA

Tianxing Chen , Kaixuan Wang , Zhaohui Yang , Yuhao Zhang , Zanxin Chen , Baijun Chen , Wanxi Dong , Ziyuan Liu
show 91 more authors
Dong Chen Tianshuo Yang Haibao Yu Xiaokang Yang Yusen Qin Zhiqiang Xie Yao Mu Ping Luo Tian Nian Weiliang Deng Yiheng Ge Yibin Liu Zixuan Li Dehui Wang Zhixuan Liang Haohui Xie Rijie Zeng Yunfei Ge Peiqing Cong Guannan He Zhaoming Han Ruocheng Yin Jingxiang Guo Lunkai Lin Tianling Xu Hongzhe Bi Xuewu Lin Tianwei Lin Shujie Luo Keyu Li Ziyan Zhao Ke Fan Heyang Xu Bo Peng Wenlong Gao Dongjiang Li Feng Jin Hui Shen Jinming Li Chaowei Cui Yu Chen Yaxin Peng Lingdong Zeng Wenlong Dong Tengfei Li Weijie Ke Jun Chen Erdemt Bao Tian Lan Tenglong Liu Jin Yang Huiping Zhuang Baozhi Jia Shuai Zhang Zhengfeng Zou Fangheng Guan Tianyi Jia Ke Zhou Hongjiu Zhang Yating Han Cheng Fang Yixian Zou Chongyang Xu Qinglun Zhang Shen Cheng Xiaohe Wang Ping Tan Haoqiang Fan Shuaicheng Liu Jiaheng Chen Chuxuan Huang Chengliang Lin Kaijun Luo Boyu Yue Yi Liu Jinyu Chen Zichang Tan Liming Deng Shuo Xu Zijian Cai Shilong Yin Hao Wang Hongshan Liu Tianyang Li Long Shi Ran Xu Huilin Xu Zhengquan Zhang Congsheng Xu Jinchang Yang Feng Xu
This is my paper · ORCID
classification cs.ROcs.AIcs.LGcs.MA
keywords bimanualmanipulationdual-armrobotsbenchmarkingpolicygeneralization3Drepresentationlearningdiffusiondeformableobjectsim-to-realtransfer
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

This paper describes a large competition arranged to answer a concrete question: when data-driven policies are trained on dual-arm robot tasks, what makes them work on new configurations? To find out, the organizers ran 17 tasks in three stages—simple simulation, domain-randomized simulation with language instructions, and physical robots—and gathered 64 teams with over 400 participants. The report argues that the winning solutions show explicit 3D representations beat 2D-only vision, that scaling data quantity only helps when data quality is also controlled, and that model capacity should align with task complexity. It also argues that success/failure metrics under-reward policies that complete most of a task but fail at the final step. If the organizers' framing is right, the competition provides a reusable testbed and concrete design recipes for robust bimanual manipulation.

What carries the argument

The evaluation protocol itself is the main instrument: three rounds, 17 tasks, 100-trial evaluation in simulation and 20 trials in the real world, with domain randomization in Round 2 and a seen/unseen split in the real-world round. Within that instrument, the decisive mechanisms that carry the argument are the two winning solution designs—AnchorDP3, a sparse-keypose diffusion policy anchored to object affordances using 3D point clouds, and SEM, a spatial enhancer that lifts 2D features into 3D embeddings and encodes the robot joint graph—because their success is the report's chief evidence for the 3D-representation insight.

What would settle it

Re-run the real-world round with all 20 trials per task on unseen backgrounds and with no high-quality demonstration release before submission, and retrain the leading simulation solution without its simulation-provided segmentation masks; if scores drop in either case, the generalization reading and the 3D-representation insight would need to be revised.

Watch

Extended reading notes

Core claim

The paper reports the design and outcome of a three-stage competition built to test whether data-driven dual-arm manipulation policies can generalize beyond their training conditions. Its central finding is that the top-performing solutions, in simulation rounds with and without domain randomization and on physical robots, consistently used explicit 3D inputs rather than pure 2D vision, while also benefiting from high-quality demonstration data and from model capacity matched to task difficulty. A second finding is that outcome-only scoring hides substantial partial competence: a team that completed almost the whole cap-insertion sequence scored nearly zero because only the final precise outcome counted.

Load-bearing premise

The benchmark's conclusions assume that its scores measure generalizable dual-arm skill, even though most real-world test trials used the exact configuration shown in the late-released demonstrations and the top 3D method received privileged simulation segmentation masks.

Editorial extensions

If this is right

  • Future dual-arm policies should embed explicit 3D structure, such as point clouds, depth-lifted features, or graph-level robot state, rather than relying on multi-view 2D images alone.
  • A two-stage data recipe that pre-trains on large medium-quality demonstrations and fine-tunes on small high-quality ones is a practical way to bridge embodiment and domain gaps.
  • Benchmarks should adopt progress-aware scoring, since binary success/failure under-credits policies that complete most of a long manipulation sequence.
  • Task-specific lightweight encoders feeding a shared action expert can prevent negative transfer in single-model multi-task deployments.
  • Model selection should be driven by task complexity, with long-horizon coordination tasks requiring higher-capacity models.

Reading between the lines

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

  • The large gap between near-ceiling simulation scores and low real-world scores suggests that sim-to-real transfer, not algorithmic ceilings, is the binding constraint; a natural next benchmark would hold the algorithm fixed and vary only simulation fidelity.
  • Because the 3D-representation insight rests on a solution that used simulator-provided segmentation masks, an equally plausible reading is that the benefit comes from privileged object masks; a testable follow-up would retrain the same architecture with masks withheld.
  • The competition format produces a public corpus of per-task success rates across dozens of independently designed policies; mining that corpus for which task attributes predict success could yield quantitative guidance beyond the report's qualitative insights.
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

3 major / 6 minor

Summary. This paper reports the RoboTwin Dual-Arm Collaboration Challenge held at the CVPR 2025 MEIS Workshop. The challenge has three stages: Simulation Round 1 (five rigid-body tasks plus one visuo-tactile task on RoboTwin 1.0), Simulation Round 2 (six rigid-body tasks with domain randomization on RoboTwin 2.0), and a Real-World Round (five tasks on the AgileX COBOT-Magic platform). The paper describes the rules, scoring, and per-task results for each round, highlights the winning solutions AnchorDP3 and SEM, and derives a set of insights about model capacity, data quantity and quality, multimodal fusion, language grounding, and data preprocessing. It closes with future directions, including progress-aware evaluation metrics.

Significance. As a competition report, the paper documents a substantial community effort: 64 teams and over 400 participants, 17 tasks across simulation and real hardware, and two well-described winning solutions with public references. The acknowledgment that binary success metrics under-reward partially successful policies (Section 4.6 and Section 7) is a useful and honest limitation statement. The main scientific value would be as a benchmark for generalizable bimanual manipulation, but that value is currently undercut by the real-world evaluation protocol and by a confounded 3D-representation insight, so the paper needs revision before the central claims can be accepted.

major comments (3)
  1. [Section 2.3.1, Table 3] The real-world round does not currently establish that the benchmark measures generalization. One week before submission, the organizers released 20 high-quality demonstrations per task collected in the official competition setting with a clean tabletop and fixed, visible backgrounds, and 15 of the 20 evaluation trials per task were conducted under this same seen configuration. A participant could therefore earn up to 75% of the real-world points by fine-tuning on the released target-domain demonstrations. The reported aggregate scores are not split by seen versus unseen background conditions, so the results are as consistent with adaptation or overfitting to the evaluation environment as with generalization. Please report per-condition success rates for each task (or at least a seen/unseen split) and state whether the released demonstrations could be used for fine-tuning; if they could, the generalization claim should be evaluated with a no-target-demo baseline or clearly qualified.
  2. [Section 3.1 and Section 3] The insight that explicit 3D representations improve sample efficiency and generalization is confounded by privileged information. AnchorDP3 'leveraged the simulation environment's complete scene knowledge to automatically generate precise point-level segmentation masks for task-critical objects.' This means its strong performance may be attributable to the segmentation masks, which are unavailable for real-world deployment, rather than to the 3D point-cloud representation itself. To support the claimed advantage of 3D representations, the paper should either report an ablation of AnchorDP3 without the simulation-provided masks, or temper the insight to state that 3D representations combined with privileged segmentation information were beneficial in simulation.
  3. [Tables 1-3 and Section 2.1.2] All reported results are point estimates without error bars, confidence intervals, or statistical tests. This matters for the 'rigorous testbed' claim: Table 3 is based on only 20 trials per task per team, so the difference between the average scores (for example, 0.30 for Fold Towel and 0.58 for Cap Pen) is within binomial noise. The paper should provide per-team per-task trial counts and, at minimum, binomial confidence intervals, and should avoid over-interpreting small numerical differences between teams or tasks.
minor comments (6)
  1. [Section 2.1.2] The sentence 'the highest success rate for all tasks exceeding 97' is ambiguous and, if read as the average success rate, inconsistent with Table 1, where average scores are as low as 51.9. If it means the maximum per-task score, it should be rephrased as such.
  2. [Section 2.3.1 and Section 2.3.2] The real-world round is first described as containing 'Five dual-arm manipulation tasks' and then as requiring a model to 'solve all six tasks'; Table 3 lists only five tasks. This inconsistency should be corrected, and the scoring formula (20 points per task, 100 total) should be checked against the corrected task count.
  3. [Section 2.1.1] The statement that 'all test-time seeds ... were kept unseen during training' is followed by the statement that background, table, and lighting conditions were kept consistent between training and evaluation; the text should clarify that only object poses and scene seeds were unseen, while the visual domain was not varied in Round 1.
  4. [Section 2.3.1] The phrase 'language instructions used during evaluation will not be provided during training' is confusing; the intended meaning is likely that the evaluation-time language instructions are held out, but this should be stated explicitly.
  5. [Section 3.1] The claims that AnchorDP3 'achieved a remarkable 98.7% success rate' and that the results in Section 3.1 describe its overall performance are not tied to any specific row of Tables 1-3; please add the task or aggregation context for this number.
  6. [Section 2.3.2 and Table 3] It is not stated whether the 'Average' column in Table 3 is over all registered teams, all valid submissions, or only the top teams shown in Figure 8; please specify the denominator and report how many teams submitted to the real-world round.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the report's claims rest on measured competition scores and disclosed protocols, not on equations or fitted parameters that reduce to their inputs.

full rationale

This is an empirical competition report, not a formal derivation. Its central claims (task difficulty, winning-solution performance, lessons about 3D representations and data quality) are supported by externally scored competition results in Tables 1–3, not by a chain of equations that presuppose the conclusions. The RoboTwin platform citations [4, 21, 22] supply the infrastructure for the benchmark, but the benchmark results are measured outcomes rather than consequences of those citations, so the self-citation is not load-bearing in a derivational sense. The winning solutions AnchorDP3 and SEM are cited as participant work and their scores were produced by the challenge evaluation, which is independently measurable; the '3D representations help' insight is an inductive interpretation of two designs, not a quantity derived from the benchmark's inputs by construction. The main validity caveat is Section 2.3.1: participants received 20 high-quality demos from the official setting one week before submission and 15 of 20 real-world trials used that seen configuration, so the real-world round partly measures fine-tuning and adaptation; Section 4.2 even states the demos 'played a pivotal role in model fine-tuning and adaptation to real-world domains.' This weakens the external-validity label 'generalizable,' but it is a disclosed evaluation-design limitation rather than a circularity of the paper's argument, and the round does reserve 5 of 20 trials per task for unseen-background testing. No specific reduction of a claimed result to its own input can be exhibited, so the circularity score is 0.

Assumptions & free parameters 0 free parameters · 1 assumptions · 0 invented entities

The report introduces no free parameters or new entities. Its conclusions rest on the domain assumption that the competition results are representative and that the selected teams' experiences capture generalizable effects.

assumptions (1)
  • domain assumption Competition scores and qualitative team reports are a reliable measure of policy generalization.
    The report treats the aggregated outcomes (Tables 1-3) and selected team experiences (Section 4) as evidence for generalizable bimanual policy learning, but no control group, error bars, or statistical tests are provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Benchmarking Generalizable Bimanual Manipulation: RoboTwin Dual-Arm Collaboration Challenge at CVPR 2025 MEIS Workshop." pith.science (2026). https://pith.science/paper/UGL3W3XL

@misc{pith2026250623351,
  author       = {Pith},
  title        = {Pith review of: Benchmarking Generalizable Bimanual Manipulation: RoboTwin Dual-Arm Collaboration Challenge at CVPR 2025 MEIS Workshop},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UGL3W3XL}},
  note         = {Machine review of arXiv:2506.23351}
}
read the original abstract

Embodied Artificial Intelligence (Embodied AI) is an emerging frontier in robotics, driven by the need for autonomous systems that can perceive, reason, and act in complex physical environments. While single-arm systems have shown strong task performance, collaborative dual-arm systems are essential for handling more intricate tasks involving rigid, deformable, and tactile-sensitive objects. To advance this goal, we launched the RoboTwin Dual-Arm Collaboration Challenge at the 2nd MEIS Workshop, CVPR 2025. Built on the RoboTwin Simulation platform (1.0 and 2.0) and the AgileX COBOT-Magic Robot platform, the competition consisted of three stages: Simulation Round 1, Simulation Round 2, and a final Real-World Round. Participants totally tackled 17 dual-arm manipulation tasks, covering rigid, deformable, and tactile-based scenarios. The challenge attracted 64 global teams and over 400 participants, producing top-performing solutions like SEM and AnchorDP3 and generating valuable insights into generalizable bimanual policy learning. This report outlines the competition setup, task design, evaluation methodology, key findings and future direction, aiming to support future research on robust and generalizable bimanual manipulation policies. The Challenge Webpage is available at https://robotwin-benchmark.github.io/cvpr-2025-challenge/.

Figures

Figures reproduced from arXiv: 2506.23351 by the authors.

Figure 1
Figure 1. Simulation Round 1 Tasks (5 Rigid Object Manipulation Tasks and 1 Tactile Manipu￾lation Task). Each task allows the submission of a single dedicated model, with a maximum score of 20 points per task (5 points for the tactile task), leading to a total score of 105 points. Notably, stage-wise scoring was employed to provide finer-grained differentiation among participants’ performance and ensure better resolution in r… view at source ↗
Figure 2
Figure 2. Distribution of Team Scores in Round 1. 103.31 102.21 98.24 97.38 93.84 91.71 89.34 87.62 87.12 82.08 60 65 70 75 80 85 90 95 100 105 110 JD-TFS TSAIL-HRL Dexmal-Bot Anyrobot momoda Digital Dynamos fightrobot Shu-robot Geely-HUST-UniTn Tartaglia Total Score [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Simulation Round 2 Tasks (6 Tasks with domain randomization). To further challenge the limits of current policy generalization and uncover more advanced solutions for dual-arm manipulation, Round 2 of the competition adopted the enhanced RoboTwin 2.0 [4] platform. Compared to Round 1, we increased the difficulty across multiple dimensions, including visual robustness, multi-task handling, and policy adaptability. Si… view at source ↗
Figures from the paper (6 more)
Figure 5
Figure 5. Figure 5: Top 10 Team Scores in Round 2. We present the average scores per task and top-10 team performances from Simulation Round 2. Each task is scored out of 100 points, with a total possible score of 600. As task difficulty increased, we observed a wider performance gap amon…
Figure 6
Figure 6. Figure 6: Real-World Round Tasks. One week prior to the final code and model submission deadline, we released an additional 20 high￾quality demonstrations per task. These were collected in the official competition setting, featuring a clean tabletop and fixed, visible background…
Figure 7
Figure 7. Figure 7: Real-World Challenge Demonstration. illustrate representative strategy executions for the tasks Pour Water and Cap Pen, respectively, as performed by participating teams. 2.3.2 Results Overview [PITH_FULL_IMAGE:figures/full_fig_p005_7.png]
Figure 8
Figure 8. Figure 8: Top 10 Team Scores in Real-World Round. We present the per-task average scores and the overall scores of the top 8 teams in the Real-World Track. Each task is scored out of 20 points, with a total possible score of 100. The results show that real-world dual-arm manipul…
Figure 9
Figure 9. Figure 9: Pipeline of AnchorDP3. placement positions. This mirrors human motor control, where conscious planning occurs only at kinematic inflection points while transit phases remain subconscious. This sparse representation offers multiple advantages: (1) Reduced prediction spa…
Figure 10
Figure 10. Figure 10: Pipeline of SEM. This design is motivated by two core insights. First, purely 2D visual encoders struggle with depth ambiguity in cluttered scenes—a limitation overcome by sampling candidate depths across views and projecting pixel features into 3D positional embeddin…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Bimanual Manipulation Within an 8 GB Budget: Zero-Copy Sensing and Quantized ACT on an Entry-Level Jetson

    cs.RO 2026-08 conditional novelty 6.0 of 10

    A bimanual ACT policy runs at 10 Hz on an 8 GB Jetson Orin Nano Super with roughly 90-95% task success, and the paper documents when quantization is necessary and which layers TensorRT refuses to quantize.

  2. RoboDojo: A Unified Sim-and-Real Benchmark for Comprehensive Evaluation of Generalist Robot Manipulation Policies

    cs.RO 2026-07 conditional novelty 6.0 of 10

    RoboDojo unifies 42 capability-oriented sim tasks and 18 multi-embodiment real tasks with shared infrastructure and a 30-policy leaderboard that shows current VLAs remain far from reliable generalist manipulation.

Reference graph

Works this paper leans on

29 extracted references · 5 canonical work pages · cited by 2 Pith papers

  1. [1]

    pi_0: A vision-language-action flow model for general robot control

    Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. pi_0: A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024

    arXiv

  2. [2]

    Rt-2: Vision-language-action models transfer web knowledge to robotic control

    Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023

    arXiv 2023

  3. [3]

    Rt-1: Robotics transformer for real-world control at scale

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

    arXiv 2022

  4. [4]

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

    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

    arXiv 2025

  5. [5]

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

    Tianxing Chen, Yao Mu, Zhixuan Liang, Zanxin Chen, Shijia Peng, Qiangyu Chen, Mingkun Xu, Ruizhen Hu, Hongyuan Zhang, Xuelong Li, et al. G3flow: Generative 3d semantic flow for pose-aware and generalizable object manipulation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 1735–1744, 2025

    2025

  6. [6]

    Diffusion policy: Visuomotor policy learning via action diffusion

    Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. The International Journal of Robotics Research, page 02783649241273668, 2023

    2023

  7. [7]

    Embodied-ai-guide, January 2025

    Lumina-Embodied-AI-Community Embodied-AI-Guide-Contributors. Embodied-ai-guide, January 2025

    work page 2025

  8. [8]

    Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation

    Zipeng Fu, Tony Z Zhao, and Chelsea Finn. Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation. arXiv preprint arXiv:2401.02117, 2024

    arXiv 2024

Show all 29 references
  1. [9]

    3d diffuser actor: Policy diffusion with 3d scene representations

    Tsung-Wei Ke, Nikolaos Gkanatsios, and Katerina Fragkiadaki. 3d diffuser actor: Policy diffusion with 3d scene representations. arXiv preprint arXiv:2402.10885, 2024

    2024 arXiv

  2. [10]

    Fine-tuning vision-language-action models: Optimizing speed and success

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

    2025 arXiv

  3. [11]

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

    Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan P Foster, Pannag R Sanketi, Quan Vuong, et al. Openvla: An open-source vision-language- action model. In 8th Annual Conference on Robot Learning

  4. [12]

    Maniskill-vitac 2025: Challenge on manipulation skill learning with vision and tactile sensing

    Chuanyu Li, Renjun Dang, Xiang Li, Zhiyuan Wu, Jing Xu, Hamidreza Kasaei, Roberto Calandra, Nathan Lepora, Shan Luo, Hao Su, et al. Maniskill-vitac 2025: Challenge on manipulation skill learning with vision and tactile sensing. arXiv preprint arXiv:2411.12503, 2024

    2025 arXiv

  5. [13]

    Cogact: A foundational vision-language-action model for synergizing cognition and action in robotic manipulation

    Qixiu Li, Yaobo Liang, Zeyu Wang, Lin Luo, Xi Chen, Mozheng Liao, Fangyun Wei, Yu Deng, Sicheng Xu, Yizhong Zhang, et al. Cogact: A foundational vision-language-action model for synergizing cognition and action in robotic manipulation. arXiv preprint arXiv:2411.19650, 2024

    2024 arXiv

  6. [14]

    Adaptdiffuser: Diffusion models as adaptive self-evolving planners

    Zhixuan Liang, Yao Mu, Mingyu Ding, Fei Ni, Masayoshi Tomizuka, and Ping Luo. Adaptdiffuser: Diffusion models as adaptive self-evolving planners. In International Conference on Machine Learning , pages 20725–20745. PMLR, 2023

    2023

  7. [15]

    Skilldiffuser: Interpretable hierarchical planning via skill abstractions in diffusion-based task execution

    Zhixuan Liang, Yao Mu, Hengbo Ma, Masayoshi Tomizuka, Mingyu Ding, and Ping Luo. Skilldiffuser: Interpretable hierarchical planning via skill abstractions in diffusion-based task execution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , ...

    2024

  8. [16]

    Dexhanddiff: Interaction-aware diffusion planning for adaptive dexterous manipulation

    Zhixuan Liang, Yao Mu, Yixiao Wang, Tianxing Chen, Wenqi Shao, Wei Zhan, Masayoshi Tomizuka, Ping Luo, and Mingyu Ding. Dexhanddiff: Interaction-aware diffusion planning for adaptive dexterous manipulation. In Proceedings of the Computer Vision and Pattern Recognition Conferen...

    2025

  9. [17]

    Sem: Enhancing spatial understanding for robust robot manipulation

    Xuewu Lin, Tianwei Lin, Lichao Huang, Hongyu Xie, Yiwei Jin, Keyu Li, and Zhizhong Su. Sem: Enhancing spatial understanding for robust robot manipulation. arXiv preprint arXiv:2505.16196, 2025

    2025

  10. [18]

    Rdt-1b: a diffusion foundation model for bimanual manipulation

    Songming Liu, Lingxuan Wu, Bangguo Li, Hengkai Tan, Huayu Chen, Zhengyi Wang, Ke Xu, Hang Su, and Jun Zhu. Rdt-1b: a diffusion foundation model for bimanual manipulation. arXiv preprint arXiv:2410.07864, 2024. 12

    2024 arXiv

  11. [19]

    Avr: Active vision-driven robotic precision manipulation with viewpoint and focal length optimization

    Yushan Liu, Shilong Mu, Xintao Chao, Zizhen Li, Yao Mu, Tianxing Chen, Shoujie Li, Chuqiao Lyu, Xiao- ping Zhang, and Wenbo Ding. Avr: Active vision-driven robotic precision manipulation with viewpoint and focal length optimization. arXiv preprint arXiv:2503.01439, 2025

    2025

  12. [20]

    Manicm: Real-time 3d diffusion policy via consistency model for robotic manipulation

    Guanxing Lu, Zifeng Gao, Tianxing Chen, Wenxun Dai, Ziwei Wang, Wenbo Ding, and Yansong Tang. Manicm: Real-time 3d diffusion policy via consistency model for robotic manipulation. arXiv preprint arXiv:2406.01586, 2024

    2024 arXiv

  13. [21]

    Robotwin: Dual-arm robot benchmark with generative digital twins

    Yao Mu, Tianxing Chen, Zanxin Chen, Shijia Peng, Zhiqian Lan, Zeyu Gao, Zhixuan Liang, Qiaojun Yu, Yude Zou, Mingkun Xu, et al. Robotwin: Dual-arm robot benchmark with generative digital twins. In Proceedings of the Computer Vision and Pattern Recognition Conference , pages 27...

    2025

  14. [22]

    Robotwin: Dual-arm robot benchmark with generative digital twins (early version)

    Yao Mu, Tianxing Chen, Shijia Peng, Zanxin Chen, Zeyu Gao, Yude Zou, Lunkai Lin, Zhiqiang Xie, and Ping Luo. Robotwin: Dual-arm robot benchmark with generative digital twins (early version). In European Conference on Computer Vision, pages 264–273. Springer, 2025

    2025

  15. [23]

    Octo: An open-source generalist robot policy

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

    2024 arXiv

  16. [24]

    Rise: 3d perception makes real-world robot imitation simple and effective

    Chenxi Wang, Hongjie Fang, Hao-Shu Fang, and Cewu Lu. Rise: 3d perception makes real-world robot imitation simple and effective. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2870–2877. IEEE, 2024

    2024

  17. [25]

    Dexvla: Vision- language model with plug-in diffusion expert for general robot control

    Junjie Wen, Yichen Zhu, Jinming Li, Zhibin Tang, Chaomin Shen, and Feifei Feng. Dexvla: Vision- language model with plug-in diffusion expert for general robot control. arXiv preprint arXiv:2502.05855, 2025

    2025 arXiv

  18. [26]

    Tinyvla: Toward fast, data-efficient vision- language-action models for robotic manipulation.IEEE Robotics and Automation Letters, 10(4):3988–3995, 2025

    Junjie Wen, Yichen Zhu, Jinming Li, Minjie Zhu, Zhibin Tang, Kun Wu, Zhiyuan Xu, Ning Liu, Ran Cheng, Chaomin Shen, Yaxin Peng, Feifei Feng, and Jian Tang. Tinyvla: Toward fast, data-efficient vision- language-action models for robotic manipulation.IEEE Robotics and Automation...

    2025

  19. [27]

    Latent action pretraining from videos

    Seonghyeon Ye, Joel Jang, Byeongguk Jeon, Se June Joo, Jianwei Yang, Baolin Peng, Ajay Mandlekar, Reuben Tan, Yu-Wei Chao, Bill Yuchen Lin, et al. Latent action pretraining from videos. InCoRL 2024 Workshop on Whole-body Control and Bimanual Manipulation: Applications in Human...

    2024

  20. [28]

    3d diffusion policy

    Yanjie Ze, Gu Zhang, Kangning Zhang, Chenyuan Hu, Muhan Wang, and Huazhe Xu. 3d diffusion policy. arXiv e-prints, pages arXiv–2403, 2024

    2024

  21. [29]

    Anchordp3: 3d affordance guided sparse diffusion policy for robotic manipulation, 2025

    Ziyan Zhao, Ke Fan, He-Yang Xu, Ning Qiao, Bo Peng, Wenlong Gao, Dongjiang Li, and Hui Shen. Anchordp3: 3d affordance guided sparse diffusion policy for robotic manipulation, 2025. 13

    2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.