pith. machine review for the scientific record. sign in

arxiv: 2605.13428 · v1 · submitted 2026-05-13 · 💻 cs.RO

Recognition: unknown

SID: Sliding into Distribution for Robust Few-Demonstration Manipulation

Yicheng Ma , Wei Yu , Zhian Su , Xidan Zhang , Huixu Dong
Authors on Pith no claims yet

Pith reviewed 2026-05-14 19:12 UTC · model grok-4.3

classification 💻 cs.RO
keywords few-shot robotic manipulationdistribution shift recoveryobject-centric motion fieldegocentric policyconditioned flow matchingreal-world robot tasksout-of-distribution generalization
0
0 comments X

The pith

A learned object-centric motion field from two demonstrations iteratively guides a robot into the reliable region of an egocentric policy for robust manipulation under large shifts.

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

The paper presents Sliding into Distribution (SID) as a structured approach that separates distribution recovery from task execution in few-demonstration robotic manipulation. From canonicalized demonstrations it learns a motion field that supplies large corrective actions when the system is far from the demonstrated manifold and naturally damps to zero near it. Once inside the reliable operating region, a lightweight egocentric policy trained via conditioned flow matching completes the task, with kinematically consistent point-cloud augmentation preserving consistency. Experiments across six real-world tasks show roughly 90 percent success from only two demonstrations even under out-of-distribution initial poses, viewpoints, distractors, and disturbances.

Core claim

SID learns an object-centric motion field from canonicalized demonstrations to iteratively slide the system toward the demonstrated manifold and into the reliable operating region of a lightweight egocentric execution policy trained with conditioned flow matching, supported by kinematically consistent point-cloud reprojection augmentation.

What carries the argument

The object-centric motion field, which generates distance-dependent corrective motions that vanish near the demonstration manifold and thereby performs online distribution recovery.

If this is right

  • Only two demonstrations suffice for approximately 90 percent success across six real-world manipulation tasks under out-of-distribution initializations.
  • Success drops by less than 10 percent when distractors and external disturbances are added.
  • The motion field enables reliable reaching, after which the egocentric policy handles precise task-specific actions.
  • Kinematically consistent point-cloud reprojection augmentation preserves action-observation consistency during policy training.

Where Pith is reading between the lines

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

  • The separation of reaching via a motion field and execution via an egocentric policy could be applied to longer-horizon tasks by learning sequential manifolds.
  • The framework may lower the data-collection burden in physical robotics by exploiting geometric structure rather than scaling demonstrations.
  • Extending the motion field to incorporate additional sensor modalities could further improve robustness without increasing demonstration count.

Load-bearing premise

Demonstrations can be canonicalized into a reliable object-centric motion field that guides the robot into the operating region of the egocentric policy despite substantial pose and viewpoint shifts.

What would settle it

Measure whether the learned motion field consistently reduces distance to the demonstration manifold in real-robot trials with novel object poses and camera viewpoints, or whether success falls substantially below 90 percent under those conditions.

Figures

Figures reproduced from arXiv: 2605.13428 by Huixu Dong, Wei Yu, Xidan Zhang, Yicheng Ma, Zhian Su.

Figure 1
Figure 1. Figure 1: SID combines an object-centric motion field with an egocentric execution policy. From a few demonstrations, the field defines a smooth descent that slides OOD states toward the demonstrated support, bringing execution back into the policy’s reliable operating region. and perturbations brittle—a failure mode frequently observed in offline imitation with low-coverage data [30]. Existing learning-based approa… view at source ↗
Figure 2
Figure 2. Figure 2: SID Overview: SID comprises three components: (a) an object-centric motion field that predicts sliding steps in object￾centric space; (b) an egocentric data augmentation module that generates augmented ID/OOD point-cloud observations via projection and reconstruction; and (c) an egocentric execution policy that predicts actions from point clouds, gripper width, and a task key, with an auxiliary ID-confiden… view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of the object-centric representation space [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Inference pipelines. (a) Open-loop switches from motion-field sliding (stopped by a field-norm threshold) to policy execution. (b) Closed-loop uses ID/pose confidence to route between policy, motion-field sliding, and recovery. At ∼ pdata(· | o seg t , ktask) and a base sample A0 t ∼ p0(·) with p0 = N (0, I). We define a linear probability path indexed by the flow time τ ∈ [0, 1]: A τ t = (1 − τ ) A 0 t + … view at source ↗
Figure 5
Figure 5. Figure 5: Visualization of the selected tasks. For each task, the robot motion is decomposed into multiple sub-steps, visualized [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Results on recomposed tasks. We report the success [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Hardware setup used in our experiments. The platform [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Stage-wise decomposition of the six main tasks, along with the target object associated with each stage. [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: ID settings (baselines). Dashed regions indicate the in-distribution (ID) object-placement areas used to collect training data for imitation baselines. OOD evaluations place the relevant object(s) outside these regions while keeping the rest of the setup unchanged. and continue the task. By contrast, the fixed-camera policy operates on a viewpoint-absolute global observation that cannot be altered by the r… view at source ↗
Figure 10
Figure 10. Figure 10: Cross-object generalization on PnP-Box. SID is trained with demonstrations using object A and evaluated on unseen test objects B–D in the PnP-Box task. The test objects are arranged from high to low visual similarity with respect to the training object A, covering variations in geometry and appearance. Task 2 demos 10 demos 50 demos Hang Tape 19/25 20/25 22/25 Hang Cup 21/25 20/25 23/25 PnP-Box 20/25 22/2… view at source ↗
Figure 11
Figure 11. Figure 11: Visualization of MAKE COFFEE. S1 S2 S3 S4 S5 S6 Overall Success / 25 24/25 20/25 20/25 16/25 14/25 14/25 14/25 TABLE IX: Per-stage success counts on the extended MAKE COFFEE task over 25 trials. S1–S6 correspond to the six stages described before. c) Results.: Table IX summarizes the quantitative results on the extended task. Despite the increased task complexity and the inclusion of more fine-grained int… view at source ↗
Figure 12
Figure 12. Figure 12: Visualization of the recomposed tasks. J. Additional Figures [PITH_FULL_IMAGE:figures/full_fig_p018_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Illustration of workspace-wide generalization with [PITH_FULL_IMAGE:figures/full_fig_p019_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Illustration of augmented trajectories for O [PITH_FULL_IMAGE:figures/full_fig_p020_14.png] view at source ↗
read the original abstract

Generalizing robotic manipulation across object poses, viewpoints, and dynamic disturbances is difficult, especially with only a few demonstrations. End-to-end visuomotor policies are expressive but data-hungry, while planning and optimization satisfy explicit constraints but do not directly capture the interaction strategies demonstrated by humans. We propose Sliding into Distribution (SID), a structured framework that learns an object-centric motion field from canonicalized demonstrations to iteratively slide the system toward the demonstrated manifold and into the reliable operating region of a lightweight egocentric execution policy, mitigating out-of-distribution (OOD) execution. The motion field provides large corrective motions when far from the demonstration manifold and naturally vanishes near convergence, enabling robust reaching under substantial pose and viewpoint shifts. Within the reached regime, an egocentric policy trained with conditioned flow matching performs task-specific manipulation, supported by kinematically consistent point-cloud reprojection augmentation that preserves action-observation consistency. Across six real-world tasks, SID achieves approximately 90% success under OOD initializations with only two demonstrations, with under a 10% drop under distractors and external disturbances. Overall, SID provides a new paradigm for few-shot manipulation: explicitly managing distribution shift via online distribution recovery.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper introduces Sliding into Distribution (SID), a hybrid framework for few-shot robotic manipulation. It learns an object-centric motion field from two canonicalized demonstrations to iteratively correct pose and viewpoint shifts, sliding the system into the operating region of a lightweight egocentric policy trained via conditioned flow matching with point-cloud reprojection augmentation. The central empirical claim is that this yields ~90% success across six real-world tasks under substantial OOD initializations, with <10% degradation under distractors and external disturbances.

Significance. If the performance claims and the motion-field guidance mechanism hold under rigorous verification, the work would offer a practical new paradigm for robust few-demonstration manipulation that explicitly manages distribution shift rather than relying solely on data scale or end-to-end generalization. The combination of a vanishing corrective field with a lightweight execution policy is conceptually attractive for real-robot deployment.

major comments (3)
  1. [§4.2] §4.2 and Eq. (3)–(5): The motion field is asserted to 'provide large corrective motions when far' and 'naturally vanish near convergence,' yet no Lyapunov-style stability argument, contraction mapping, or convergence-rate analysis is supplied to show that the field remains well-behaved when canonicalization errors arise from viewpoint shifts or when external disturbances act on the real dynamics.
  2. [Table 1] Table 1 and §5.3: Success rates are reported as point estimates (~90% and <10% drop) without per-task standard deviations, number of trials, or statistical significance tests against the strongest baselines; this makes it impossible to judge whether the headline performance difference is reliable or could be explained by run-to-run variance.
  3. [§5.4] §5.4 (ablation study): The paper does not report an ablation that isolates the contribution of the learned motion field versus the egocentric policy alone under the same OOD pose/viewpoint conditions; without this, it is unclear whether the iterative guidance is load-bearing for the reported robustness.
minor comments (2)
  1. [Figure 3] Figure 3: The visualization of the motion field on point clouds would benefit from an explicit color scale or vector-length legend so readers can verify the claimed 'large corrective' to 'vanishing' behavior.
  2. [§3.1] §3.1: The precise procedure for obtaining the canonical frame from a single demonstration is described at a high level; a short pseudocode block or explicit transformation equations would improve reproducibility.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive feedback on our manuscript. We address each major comment below and describe the corresponding revisions.

read point-by-point responses

  1. Referee: [§4.2] §4.2 and Eq. (3)–(5): The motion field is asserted to 'provide large corrective motions when far' and 'naturally vanish near convergence,' yet no Lyapunov-style stability argument, contraction mapping, or convergence-rate analysis is supplied to show that the field remains well-behaved when canonicalization errors arise from viewpoint shifts or when external disturbances act on the real dynamics.

    Authors: We acknowledge that a formal stability analysis is absent. The motion field is designed from canonicalized demonstrations to produce large corrections far from the manifold and to vanish near convergence, with this behavior empirically confirmed across six real-world tasks under viewpoint shifts and disturbances. In the revision we will add a dedicated discussion of the empirical convergence behavior, including plots of field magnitude versus distance to the manifold, while noting that a full Lyapunov proof remains future work. revision: partial

  2. Referee: [Table 1] Table 1 and §5.3: Success rates are reported as point estimates (~90% and <10% drop) without per-task standard deviations, number of trials, or statistical significance tests against the strongest baselines; this makes it impossible to judge whether the headline performance difference is reliable or could be explained by run-to-run variance.

    Authors: We agree that statistical details are required for assessing reliability. The revised manuscript will expand Table 1 to report the number of trials per task, per-task standard deviations, and results of statistical significance tests against the strongest baselines. revision: yes

  3. Referee: [§5.4] §5.4 (ablation study): The paper does not report an ablation that isolates the contribution of the learned motion field versus the egocentric policy alone under the same OOD pose/viewpoint conditions; without this, it is unclear whether the iterative guidance is load-bearing for the reported robustness.

    Authors: We recognize the importance of isolating the motion field's contribution. The revised manuscript will include a new ablation that evaluates the egocentric policy alone under identical OOD pose and viewpoint conditions, directly comparing it to the full SID framework to quantify the benefit of the iterative guidance. revision: yes

Circularity Check

0 steps flagged

No circularity in derivation chain

full rationale

The paper presents SID as a framework that learns an object-centric motion field from canonicalized demonstrations, with its corrective behavior described as emerging from the learned representation rather than predefined by the target success metric. No equations or claims in the provided text reduce by construction to fitted parameters, self-referential definitions, or self-citation chains. The central empirical claims (approximately 90% success on six tasks with two demonstrations) rest on real-world experimental results under OOD conditions, not on tautological derivations. The framework is self-contained against external benchmarks with independent validation.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the assumption that few demonstrations suffice to define a usable manifold and that the learned motion field reliably transitions control to the egocentric policy; no explicit free parameters or invented physical entities are detailed in the abstract.

axioms (1)
  • domain assumption Few demonstrations can be canonicalized to learn an effective object-centric motion field for guiding the system into the egocentric policy's operating region
    This is the load-bearing premise enabling the 90% success claim with only two demonstrations under OOD conditions.
invented entities (1)
  • object-centric motion field no independent evidence
    purpose: To provide large corrective motions when far from the demonstration manifold that naturally vanish near convergence
    Introduced as the core mechanism for online distribution recovery in SID.

pith-pipeline@v0.9.0 · 5515 in / 1336 out tokens · 58691 ms · 2026-05-14T19:12:49.481850+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

65 extracted references · 44 canonical work pages · 7 internal anchors

  1. [1]

    Collins, Mrinal Jain, and Animesh Garg

    Ezra Ameperosa, Jeremy A. Collins, Mrinal Jain, and Animesh Garg. Rocoda: Counterfactual data augmen- tation for data-efficient robot learning from demonstra- tions, 2025. URL https://arxiv.org/abs/2411.16959

    work page arXiv 2025

  2. [2]

    SAM 3: Segment Anything with Concepts

    Nicolas Carion, Laura Gustafson, Yuan-Ting Hu, Shoub- hik Debnath, Ronghang Hu, Didac Suris, Chaitanya Ryali, Kalyan Vasudev Alwala, Haitham Khedr, An- drew Huang, Jie Lei, Tengyu Ma, Baishan Guo, Arpit Kalla, Markus Marks, Joseph Greer, Meng Wang, Peize Sun, Roman R ¨adle, Triantafyllos Afouras, Effrosyni Mavroudi, Katherine Xu, Tsung-Han Wu, Yu Zhou, Li...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  3. [3]

    Object-centric representations improve policy generalization in robot manipulation

    Alexandre Chapin, Bruno Machado, Emmanuel Dellan- drea, and Liming Chen. Object-centric representations improve policy generalization in robot manipulation. arXiv preprint arXiv:2505.11563, 2025

    work page arXiv 2025

  4. [4]

    Genaug: Retargeting behaviors to unseen situations via gener- ative augmentation,

    Zoey Chen, Sho Kiami, Abhishek Gupta, and Vikash Kumar. Genaug: Retargeting behaviors to unseen situ- ations via generative augmentation, 2023. URL https: //arxiv.org/abs/2302.06671

    work page arXiv 2023

  5. [5]

    Semantically con- trollable augmentations for generalizable robot learn- ing,

    Zoey Chen, Zhao Mandi, Homanga Bharadhwaj, Mohit Sharma, Shuran Song, Abhishek Gupta, and Vikash Kumar. Semantically controllable augmentations for generalizable robot learning, 2024. URL https://arxiv. org/abs/2409.00951

    work page arXiv 2024

  6. [6]

    Diffusion policy: Visuomotor policy learning via action diffusion, 2024

    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, 2024. URL https://arxiv.org/abs/2303. 04137

    2024

  7. [7]

    Learning a thousand tasks in a day.Science Robotics, 10(108), November 2025

    Kamil Dreczkowski, Pietro Vitiello, Vitalis V osylius, and Edward Johns. Learning a thousand tasks in a day.Science Robotics, 10(108), November 2025. ISSN 2470-9476. doi: 10.1126/scirobotics.adv7594. URL http://dx.doi.org/10.1126/scirobotics.adv7594

    work page doi:10.1126/scirobotics.adv7594 2025

  8. [8]

    Behavior retrieval: Few-shot imitation learning by querying unlabeled datasets, 2023

    Maximilian Du, Suraj Nair, Dorsa Sadigh, and Chelsea Finn. Behavior retrieval: Few-shot imitation learning by querying unlabeled datasets, 2023. URL https://arxiv.org/ abs/2304.08742

    work page arXiv 2023

  9. [9]

    One-Shot Imitation Learning

    Yan Duan, Marcin Andrychowicz, Bradly C. Stadie, Jonathan Ho, Jonas Schneider, Ilya Sutskever, Pieter Abbeel, and Wojciech Zaremba. One-shot imitation learning, 2017. URL https://arxiv.org/abs/1703.07326

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  10. [10]

    Kalm: Keypoint abstraction using large models for object-relative imitation learning

    Xiaolin Fang, Bo-Ruei Huang, Jiayuan Mao, Jasmine Shone, Joshua B Tenenbaum, Tom ´as Lozano-P ´erez, and Leslie Pack Kaelbling. Kalm: Keypoint abstraction using large models for object-relative imitation learning. In 2025 IEEE International Conference on Robotics and Automation (ICRA), pages 8307–8314. IEEE, 2025

    2025

  11. [11]

    One-shot visual imitation learning via meta-learning, 2017

    Chelsea Finn, Tianhe Yu, Tianhao Zhang, Pieter Abbeel, and Sergey Levine. One-shot visual imitation learning via meta-learning, 2017. URL https://arxiv.org/abs/1709. 04905

    2017

  12. [12]

    G ´omez Rodr´ıguez, Jos´e M

    Jianfeng Gao, Zhi Tao, No ´emie Jaquier, and Tamim Asfour. K-vil: Keypoints-based visual imitation learn- ing.IEEE Transactions on Robotics, 39(5):3888– 3908, October 2023. ISSN 1941-0468. doi: 10.1109/ tro.2023.3286074. URL http://dx.doi.org/10.1109/TRO. 2023.3286074

    work page doi:10.1109/tro 2023

  13. [13]

    Partmanip: Learning cross-category generalizable part manipulation policy from point cloud observations

    Haoran Geng, Ziming Li, Yiran Geng, Jiayi Chen, Hao Dong, and He Wang. Partmanip: Learning cross-category generalizable part manipulation policy from point cloud observations. InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, pages 2978–2988, 2023

    2023

  14. [14]

    Rvt: Robotic view transformer for 3d object manipulation

    Ankit Goyal, Jie Xu, Yijie Guo, Valts Blukis, Yu-Wei Chao, and Dieter Fox. Rvt: Robotic view transformer for 3d object manipulation. InConference on Robot Learning, pages 694–710. PMLR, 2023

    2023

  15. [15]

    Adaflow: Imitation learning with variance-adaptive flow-based policies, 2024

    Xixi Hu, Bo Liu, Xingchao Liu, and Qiang Liu. Adaflow: Imitation learning with variance-adaptive flow-based policies, 2024. URL https://arxiv.org/abs/2402.04292

    work page arXiv 2024

  16. [16]

    Physical Intelligence, Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, Adnan Es- mail, Michael Equi, Chelsea Finn, Niccolo Fusai, Manuel Y . Galliker, Dibya Ghosh, Lachy Groom, Karol Hausman, Brian Ichter, Szymon Jakubczak, Tim Jones, Liyiming Ke, Devin LeBlanc, Sergey Levine, Adrian Li-Bell, Mohith Mothukuri, Suraj Nair, Karl Pert...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  17. [17]

    Coarse-to-fine q-attention: Efficient learning for visual robotic manipulation via discretisa- tion

    Stephen James, Kentaro Wada, Tristan Laidlow, and Andrew J Davison. Coarse-to-fine q-attention: Efficient learning for visual robotic manipulation via discretisa- tion. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13739– 13748, 2022

    2022

  18. [18]

    Bc-z: Zero-shot task generalization with robotic imitation learning

    Eric Jang, Alex Irpan, Mohi Khansari, Daniel Kappler, Frederik Ebert, Corey Lynch, Sergey Levine, and Chelsea Finn. Bc-z: Zero-shot task generalization with robotic imitation learning. InConference on Robot Learning, pages 991–1002. PMLR, 2022

    2022

  19. [19]

    Roboexp: Action-conditioned scene graph via interactive exploration for robotic manipulation.arXiv preprint arXiv:2402.15487, 2024

    Hanxiao Jiang, Binghao Huang, Ruihai Wu, Zhuoran Li, Shubham Garg, Hooshang Nayyeri, Shenlong Wang, and Yunzhu Li. Roboexp: Action-conditioned scene graph via interactive exploration for robotic manipulation.arXiv preprint arXiv:2402.15487, 2024

    work page arXiv 2024

  20. [20]

    Coarse-to-fine imitation learning: Robot manipulation from a single demonstration

    Edward Johns. Coarse-to-fine imitation learning: Robot manipulation from a single demonstration. In2021 IEEE international conference on robotics and automation (ICRA), pages 4613–4619. IEEE, 2021

    2021

  21. [21]

    3d diffuser actor: Policy diffusion with 3d scene representations, 2024

    Tsung-Wei Ke, Nikolaos Gkanatsios, and Katerina Fragkiadaki. 3d diffuser actor: Policy diffusion with 3d scene representations, 2024. URL https://arxiv.org/abs/ 2402.10885

    work page arXiv 2024

  22. [22]

    Weijie Kong, Zhaohui Lin, Wei Yu, Haotian Guo, Zhian Su, and Huixu Dong. Affpose: An integrated rgb- based framework for simultaneous pose estimation and affordance detection in robotic tool manipulation.IEEE Robotics and Automation Letters, 10(10):10170–10177,

  23. [23]

    doi: 10.1109/LRA.2025.3598984

    work page doi:10.1109/lra.2025.3598984 2025

  24. [24]

    End-to-end training of deep visuomotor policies,

    Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. End-to-end training of deep visuomotor policies,

  25. [25]

    URL https://arxiv.org/abs/1504.00702

    work page internal anchor Pith review Pith/arXiv arXiv

  26. [26]

    Learning Hand-Eye Coordination for Robotic Grasping with Deep Learning and Large-Scale Data Collection

    Sergey Levine, Peter Pastor, Alex Krizhevsky, and Deirdre Quillen. Learning hand-eye coordination for robotic grasping with deep learning and large-scale data collection, 2016. URL https://arxiv.org/abs/1603.02199

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  27. [27]

    Language-guided object-centric diffusion policy for generalizable and collision-aware manipulation

    Hang Li, Qian Feng, Zhi Zheng, Jianxiang Feng, Zhaopeng Chen, and Alois Knoll. Language-guided object-centric diffusion policy for generalizable and collision-aware manipulation. In2025 IEEE Interna- tional Conference on Robotics and Automation (ICRA), pages 12834–12841. IEEE, 2025

    2025

  28. [28]

    Bridgevla: Input-output alignment for efficient 3d manipulation learning with vision-language models

    Peiyan Li, Yixiang Chen, Hongtao Wu, Xiao Ma, Xiang- nan Wu, Yan Huang, Liang Wang, Tao Kong, and Tieniu Tan. Bridgevla: Input-output alignment for efficient 3d manipulation learning with vision-language models. arXiv preprint arXiv:2506.07961, 2025

    work page arXiv 2025

  29. [29]

    Unidoormanip: Learning universal door manipulation policy over large- scale and diverse door manipulation environments.arXiv preprint arXiv:2403.02604, 2024

    Yu Li, Xiaojie Zhang, Ruihai Wu, Zilong Zhang, Yiran Geng, Hao Dong, and Zhaofeng He. Unidoormanip: Learning universal door manipulation policy over large- scale and diverse door manipulation environments.arXiv preprint arXiv:2403.02604, 2024

    work page arXiv 2024

  30. [30]

    A coarse- to-fine multimodal detection framework based on deep learning for robotic coating tasks.IEEE/ASME Trans- actions on Mechatronics, 31(1):639–650, 2026

    Zhaohui Lin, Haonan Dong, Weijie Kong, Haoran Huang, I-Ming Chen, and Huixu Dong. A coarse- to-fine multimodal detection framework based on deep learning for robotic coating tasks.IEEE/ASME Trans- actions on Mechatronics, 31(1):639–650, 2026. doi: 10.1109/TMECH.2025.3595263

    work page doi:10.1109/tmech.2025.3595263 2026

  31. [31]

    Learning to generalize across long-horizon tasks from human demonstrations

    Ajay Mandlekar, Danfei Xu, Roberto Mart ´ın-Mart´ın, Silvio Savarese, and Li Fei-Fei. Learning to generalize across long-horizon tasks from human demonstrations. arXiv preprint arXiv:2003.06085, 2020

    work page arXiv 2003

  32. [32]

    What Matters in Learning from Offline Human Demonstrations for Robot Manipulation

    Ajay Mandlekar, Danfei Xu, Josiah Wong, Soroush Nasiriany, Chen Wang, Rohun Kulkarni, Li Fei-Fei, Silvio Savarese, Yuke Zhu, and Roberto Mart ´ın-Mart´ın. What matters in learning from offline human demon- strations for robot manipulation.arXiv preprint arXiv:2108.03298, 2021

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  33. [33]

    kpam: Keypoint affordances for category-level robotic manipulation, 2019

    Lucas Manuelli, Wei Gao, Peter Florence, and Russ Tedrake. kpam: Keypoint affordances for category-level robotic manipulation, 2019. URL https://arxiv.org/abs/ 1903.06684

    work page arXiv 2019

  34. [34]

    Calvin: A benchmark for language- conditioned policy learning for long-horizon robot ma- nipulation tasks, 2022

    Oier Mees, Lukas Hermann, Erick Rosete-Beas, and Wolfram Burgard. Calvin: A benchmark for language- conditioned policy learning for long-horizon robot ma- nipulation tasks, 2022. URL https://arxiv.org/abs/2112. 03227

    2022

  35. [35]

    Two by two: Learning multi- task pairwise objects assembly for generalizable robot manipulation

    Yu Qi, Yuanchen Ju, Tianming Wei, Chi Chu, Lawson LS Wong, and Huazhe Xu. Two by two: Learning multi- task pairwise objects assembly for generalizable robot manipulation. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 17383–17393, 2025

    2025

  36. [36]

    Task-oriented hierarchical object decomposition for visuomotor control.arXiv preprint arXiv:2411.01284, 2024

    Jianing Qian, Yunshuang Li, Bernadette Bucher, and Dinesh Jayaraman. Task-oriented hierarchical object decomposition for visuomotor control.arXiv preprint arXiv:2411.01284, 2024

    work page arXiv 2024

  37. [37]

    Behavior transformers: Cloning k modes with one stone

    Nur Muhammad Mahi Shafiullah, Zichen Jeff Cui, Ar- iuntuya Altanzaya, and Lerrel Pinto. Behavior trans- formers: Cloningkmodes with one stone, 2022. URL https://arxiv.org/abs/2206.11251

    work page arXiv 2022

  38. [38]

    Cliport: What and where pathways for robotic manipulation,

    Mohit Shridhar, Lucas Manuelli, and Dieter Fox. Cliport: What and where pathways for robotic manipulation,

  39. [39]

    URL https://arxiv.org/abs/2109.12098

    work page arXiv

  40. [40]

    Perceiver-actor: A multi-task transformer for robotic ma- nipulation

    Mohit Shridhar, Lucas Manuelli, and Dieter Fox. Perceiver-actor: A multi-task transformer for robotic ma- nipulation. InConference on Robot Learning, pages 785–

  41. [41]

    Fast flow-based visuomotor policies via conditional optimal transport couplings, 2025

    Andreas Sochopoulos, Nikolay Malkin, Nikolaos Tsagkas, Jo ˜ao Moura, Michael Gienger, and Sethu Vijayakumar. Fast flow-based visuomotor policies via conditional optimal transport couplings, 2025. URL https://arxiv.org/abs/2505.01179

    work page arXiv 2025

  42. [42]

    Construction of bin-picking system for logistic applica- tion: A hybrid robotic gripper and vision-based grasp planning.IEEE Robotics and Automation Letters, 10(8): 8300–8307, 2025

    Zhian Su, Yicheng Ma, Haotian Guo, and Huixu Dong. Construction of bin-picking system for logistic applica- tion: A hybrid robotic gripper and vision-based grasp planning.IEEE Robotics and Automation Letters, 10(8): 8300–8307, 2025. doi: 10.1109/LRA.2025.3585393

    work page doi:10.1109/lra.2025.3585393 2025

  43. [43]

    Kite: Keypoint-conditioned policies for semantic manipulation.arXiv preprint arXiv:2306.16605, 2023

    Priya Sundaresan, Suneel Belkhale, Dorsa Sadigh, and Jeannette Bohg. Kite: Keypoint-conditioned policies for semantic manipulation.arXiv preprint arXiv:2306.16605, 2023

    work page arXiv 2023

  44. [44]

    Functo: Function-centric one-shot imi- tation learning for tool manipulation.arXiv preprint arXiv:2502.11744, 2025

    Chao Tang, Anxing Xiao, Yuhong Deng, Tianrun Hu, Wenlong Dong, Hanbo Zhang, David Hsu, and Hong Zhang. Functo: Function-centric one-shot imi- tation learning for tool manipulation.arXiv preprint arXiv:2502.11744, 2025

    work page arXiv 2025

  45. [45]

    Mimicplay: Long- horizon imitation learning by watching human play,

    Chen Wang, Linxi Fan, Jiankai Sun, Ruohan Zhang, Li Fei-Fei, Danfei Xu, Yuke Zhu, and Anima Anand- kumar. Mimicplay: Long-horizon imitation learning by watching human play.arXiv preprint arXiv:2302.12422, 2023

    work page arXiv 2023

  46. [46]

    Rise: 3d perception makes real-world robot imitation simple and effective, 2024

    Chenxi Wang, Hongjie Fang, Hao-Shu Fang, and Cewu Lu. Rise: 3d perception makes real-world robot imitation simple and effective, 2024. URL https://arxiv.org/abs/ 2404.12281

    work page arXiv 2024

  47. [47]

    Equiv- ariant diffusion policy, 2024

    Dian Wang, Stephen Hart, David Surovik, Tarik Ke- lestemur, Haojie Huang, Haibo Zhao, Mark Yeatman, Jiuguang Wang, Robin Walters, and Robert Platt. Equiv- ariant diffusion policy, 2024. URL https://arxiv.org/abs/ 2407.01812

    work page arXiv 2024

  48. [48]

    Skil: Semantic keypoint imitation learn- ing for generalizable data-efficient manipulation, 2025

    Shengjie Wang, Jiacheng You, Yihang Hu, Jiongye Li, and Yang Gao. Skil: Semantic keypoint imitation learn- ing for generalizable data-efficient manipulation, 2025. URL https://arxiv.org/abs/2501.14400

    work page arXiv 2025

  49. [49]

    Foundationpose: Unified 6d pose estimation and tracking of novel objects, 2024

    Bowen Wen, Wei Yang, Jan Kautz, and Stan Birchfield. Foundationpose: Unified 6d pose estimation and tracking of novel objects, 2024. URL https://arxiv.org/abs/2312. 08344

    2024

  50. [50]

    Cage: Causal attention enables data-efficient generalizable robotic manipulation, 2024

    Shangning Xia, Hongjie Fang, Cewu Lu, and Hao- Shu Fang. Cage: Causal attention enables data-efficient generalizable robotic manipulation, 2024. URL https: //arxiv.org/abs/2410.14974

    work page arXiv 2024

  51. [51]

    Demogen: Syn- thetic demonstration generation for data-efficient visuo- motor policy learning.arXiv preprint arXiv:2502.16932, 2025

    Zhengrong Xue, Shuying Deng, Zhenyang Chen, Yixuan Wang, Zhecheng Yuan, and Huazhe Xu. Demogen: Synthetic demonstration generation for data-efficient vi- suomotor policy learning, 2025. URL https://arxiv.org/ abs/2502.16932

    work page arXiv 2025

  52. [52]

    Equibot: Sim(3)-equivariant diffusion policy for generalizable and data efficient learning, 2024

    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, 2024. URL https://arxiv.org/abs/ 2407.01479

    work page arXiv 2024

  53. [53]

    Equiv- act: Sim(3)-equivariant visuomotor policies beyond rigid object manipulation, 2024

    Jingyun Yang, Congyue Deng, Jimmy Wu, Rika Antonova, Leonidas Guibas, and Jeannette Bohg. Equiv- act: Sim(3)-equivariant visuomotor policies beyond rigid object manipulation, 2024. URL https://arxiv.org/abs/ 2310.16050

    work page arXiv 2024

  54. [54]

    3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations, 2024

    Yanjie Ze, Gu Zhang, Kangning Zhang, Chenyuan Hu, Muhan Wang, and Huazhe Xu. 3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations, 2024. URL https://arxiv.org/abs/2403. 03954

    2024

  55. [55]

    Transporter networks: Rearranging the visual world for robotic manipulation,

    Andy Zeng, Pete Florence, Jonathan Tompson, Stefan Welker, Jonathan Chien, Maria Attarian, Travis Arm- strong, Ivan Krasin, Dan Duong, Ayzaan Wahid, Vikas Sindhwani, and Johnny Lee. Transporter networks: Rearranging the visual world for robotic manipulation,

  56. [56]

    URL https://arxiv.org/abs/2010.14406

    work page arXiv 2010

  57. [57]

    Flowpolicy: Enabling fast and robust 3d flow-based policy via consistency flow matching for robot manipulation.arXiv preprint arXiv:2412.04987, 2024

    Qinglun Zhang, Zhen Liu, Haoqiang Fan, Guanghui Liu, Bing Zeng, and Shuaicheng Liu. Flowpolicy: Enabling fast and robust 3d flow-based policy via consistency flow matching for robot manipulation, 2024. URL https://arxiv.org/abs/2412.04987

    work page arXiv 2024

  58. [58]

    One-shot imita- tion learning with invariance matching for robotic ma- nipulation, 2024

    Xinyu Zhang and Abdeslam Boularias. One-shot imita- tion learning with invariance matching for robotic ma- nipulation, 2024. URL https://arxiv.org/abs/2405.13178

    work page arXiv 2024

  59. [59]

    General- izable hierarchical skill learning via object-centric repre- sentation, 2025

    Haibo Zhao, Yu Qi, Boce Hu, Yizhe Zhu, Ziyan Chen, Heng Tian, Xupeng Zhu, Owen Howell, Haojie Huang, Robin Walters, Dian Wang, and Robert Platt. General- izable hierarchical skill learning via object-centric repre- sentation, 2025. URL https://arxiv.org/abs/2510.21121

    work page arXiv 2025

  60. [60]

    Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware

    Tony Z. Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manip- ulation with low-cost hardware, 2023. URL https://arxiv. org/abs/2304.13705

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  61. [61]

    You only teach once: Learn one-shot bimanual robotic manipulation from video demonstrations, 2025

    Huayi Zhou, Ruixiang Wang, Yunxin Tai, Yueci Deng, Guiliang Liu, and Kui Jia. You only teach once: Learn one-shot bimanual robotic manipulation from video demonstrations, 2025. URL https://arxiv.org/abs/ 2501.14208

    work page arXiv 2025

  62. [62]

    Learning generalizable manipulation policies with object-centric 3d representations.arXiv preprint arXiv:2310.14386, 2023

    Yifeng Zhu, Zhenyu Jiang, Peter Stone, and Yuke Zhu. Learning generalizable manipulation policies with object-centric 3d representations.arXiv preprint arXiv:2310.14386, 2023. APPENDIXA A. Hardware Setup Fig. 7: Hardware setup used in our experiments. The platform consists of a UR5e arm with a Robotiq 2F-85 gripper and an Intel RealSense L515 RGB-D camera...

    work page arXiv 2023

  63. [63]

    capture range

    Ablation 1: Effect of Egocentric Data Augmentation: a) Setting.:We compare two training settings: (i)w/ ego- aug, where we apply egocentric data augmentation to expand the dataset from 2 demonstrations to 50 demonstrations, and (ii)w/o ego-aug, where we train using only the original 2 demonstrations without augmentation. All other training and inference c...

  64. [64]

    Ablation 2: Egocentric vs. Fixed Camera Viewpoint: a) Setting.:To isolate the effect of observation view- point, we collect demonstrations with the egocentric camera and an external fixed camera mounted simultaneously. This yields paired observations for each timestep from identical trajectories, ensuring that the training data are matched and the only di...

  65. [65]

    in-distribution

    Summary:Overall, the ablations highlight the impor- tance of egocentric design for robust multi-stage manipulation. Egocentric data augmentation substantially improves perfor- mance, particularly on long-horizon tasks, by broadening the training support and increasing tolerance to motion-field pose estimation errors and observation drift. Moreover, egocen...