Pith. sign in

REVIEW 3 major objections 4 minor 57 references

Dynamics-Aware Meta-Imitation for Generalization to Unseen Robotic Manipulation

T0 review · 3 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read A meta-learned 3D diffusion policy generalizes to unseen manipulation tasks using one reference demonstration and a few fine-tuning steps, the paper argues.

desk verdict A well-engineered MAML + 3D diffusion system whose 'unseen task' evaluation leaks the target task through a same-task reference demo; the architecture is worth refereeing but the generalization claim needs a protocol re-run. read the letter →

arxiv 2607.15880 v1 pith:PNXAXU6S submitted 2026-07-17 cs.RO cs.LG

Zhenduo Shang , Xiyao Liu , Bohan Li , Xudong Wang , Teng Ren , Lianqing Liu , Zhi Han This is my paper
classification cs.ROcs.LG
keywords few-shotimitationlearningmeta-learning3Ddiffusionpolicyroboticmanipulationtaskgeneralizationvisual-motortrajectorycross-modalfusiondynamics-awareconditioning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that imitation-learning robots fail on new manipulation tasks because they memorize static visual cues instead of the dynamics of how the scene evolves under the expert's actions. DAMI (Dynamics-Aware Meta-Imitation) couples a full model-agnostic meta-learning objective with a 3D diffusion policy conditioned on a complete reference demonstration, a language instruction, and the current observation. The authors claim that after meta-training on base tasks, a shared initialization can be quickly fine-tuned on a few demonstrations of an unseen task, beating strong baselines on Meta-World, RLBench FS25, and a real UR10e setup. A sympathetic reader would care because this is a concrete step toward robots that adapt to new manipulation tasks without collecting thousands of demonstrations.

What carries the argument

The core machinery is the full model-agnostic meta-learning (MAML) procedure wrapped around a 3D diffusion policy with three conditioning components: VMT encodes the reference demonstration jointly from visual and motor channels with positional embeddings; U2T is a Transformer block that fuses the demonstration tokens, a frozen CLIP text feature, and the observation–timestep condition at the first downsampling layer; and TCFM uses FiLM-style modulation so every U-Net block conditions on the fused task context. The inner loop takes one action-prediction gradient step on support data, and the outer loop averages full second-order query-set gradients across tasks to update a shared initializati

What would settle it

Run the Meta-World ML10 novel-task evaluation with strict episode-level separation: sample the reference demonstration from one episode and the support/query fine-tuning windows from different episodes of the same task; if the success rate drops to the level of the no-meta-learning baselines, the central generalization claim is false. Additionally, rerun with a reference demonstration from a different but semantically related task; if the model fails to adapt, the method is copying task identity rather than dynamics.

Watch

Extended reading notes

Core claim

The central claim is that coupling a full-MAML meta-learning objective with a 3D diffusion policy conditioned on a complete reference demonstration produces an initialization that captures transferable task dynamics rather than static appearance. The Visual-Motor Trajectory (VMT) module jointly encodes visual frames and motor actions with temporal position information, letting the policy represent 'what changes under what action.' The Unpaired Unified Task (U2T) block fuses demonstration, text, and observation–timestep conditions at the first U-Net downsampling layer, and Task-Conditioned Feature Modulation (TCFM) injects the fused condition into low-level 3D features. The paper reports that

Load-bearing premise

The protocol assumes a task is 'unseen' even though the policy is given a complete expert demonstration of that exact task at inference time, and the code does not enforce episode-level separation, so the reference demonstration may also appear inside the support or query windows used for fine-tuning.

Editorial extensions

If this is right

  • If the claim holds, a robot trained once on a set of manipulation tasks can be deployed to a new task with one complete demonstration plus a short fine-tuning run, cutting data-collection cost.
  • The separation of motion-centric demonstration encoding from static appearance implies that gains should transfer to novel objects with the same task semantics, not just the exact objects seen at training.
  • Strong performance on both base direct inference and novel-task adaptation indicates that meta-learning does not sacrifice base-task proficiency, a common failure mode of prior few-shot methods.
  • DAMI outperforms MAML-enhanced baselines, suggesting the conditioning modules themselves—not the meta-objective alone—carry the generalization benefit.
  • The real-world adaptation to two novel tasks in under two minutes of compute suggests practical deployability in settings where new tasks appear frequently.

Reading between the lines

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

  • A direct test of the core claim would enforce strict episode-level separation, ensuring the reference demonstration never contributes support or query windows; if gains vanish, the generalization claim reduces to copying from the same trajectory.
  • A testable extension is to provide a reference demonstration from a different but semantically related task, or a partial demonstration, to see whether the policy extrapolates beyond the exact task it is adapting to.
  • Ablating the motor channel inside VMT could reveal whether the representation is genuinely dynamics-aware by checking whether performance degrades more on action-semantic tasks (push versus pull versus press) than on purely geometric ones.
  • The modest reported computational overhead (39.3 s versus 32.2 s per 200 adaptation steps) suggests the main deployment obstacle is the need for task-specific fine-tuning and sensitivity to reference-demonstration quality, not raw compute.
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 / 4 minor

Summary. The paper proposes DAMI, a meta-imitation framework for robotic manipulation that conditions a 3D diffusion policy on a complete reference demonstration, a text instruction, and the current observation. The method uses three new components: VMT to encode visual-motor trajectory tokens, U2T for transformer-based fusion of demo/text/observation-timestep conditions, and TCFM for FiLM-style modulation of U-Net features. Training uses full MAML with support/query batches, and novel-task adaptation fine-tunes the EMA pretrained policy on the target task's own dataset. Experiments on Meta-World ML10/ML45, RLBench FS25, and a UR10e real-world setup report large gains over DP3, Mamba, FreqPolicy, and FlowPolicy for base-task direct inference and few-shot novel-task adaptation.

Significance. If the reported gains were robust, the architecture would be a useful step toward data-efficient cross-task imitation. The paper explicitly addresses a real limitation of prior IL methods and provides a fairly complete description of the meta-training and adaptation algorithms. The inclusion of a stated limitation about episode-level separation (Sec. III-F) is transparent, but it exposes a load-bearing weakness: the 'unseen task' evaluation is contaminated by same-task reference demonstrations. The central generalization claim is therefore not yet substantiated. The method is plausible, but the experiments as reported do not isolate whether DAMI learns transferable task dynamics or simply retrieves/copies action patterns from a provided expert demo of the exact target task.

major comments (3)
  1. [Sec. III-A and Algorithm 2; Sec. III-F] The 'unseen task' evaluation leaks the target task through the reference demonstration. Sec. III-A states that reference demo τi is randomly selected from the same task dataset DTi as support/query windows, and Sec. III-F explicitly admits that 'the code does not enforce episode-level separation, so the selected demonstration episode may also contribute support or query windows.' Algorithm 2 repeats this for novel tasks: the reference trajectory d and fine-tuning batches are both drawn from the same target-task dataset D. Because VMT encodes the demo's visual states and actions and U2T injects those tokens into the denoiser, the policy can implement a retrieval policy: match current observation to the demo and copy recorded actions. The abstract's claim that DAMI learns 'underlying task logic rather than memorizing static cues' is therefore not supported by the protocol. The paper should
  2. [Tables II, III, IV; Sec. IV-B] All reported success rates are means over 3 seeds (Sec. IV-A) but no error bars, standard deviations, or significance tests are provided. This matters because several headline claims are based on small or tied margins. In Table III, DAMI ties DP3 on ML45 novel tasks at 37.53%, so the claim that DAMI 'outperforms' baselines on novel tasks is not generally true; the ML45 advantage appears only on base tasks. Table IV reports only the row 'Avg.' for RLBench FS25 without per-task results, making the 10.80% average unverifiable. The paper should report per-seed results, variance, and, where possible, statistical tests (e.g., paired bootstrap across seeds) for the main comparisons.
  3. [Sec. IV-A; Table VIII; Sec. IV-B.1] The comparison against DP3/Mamba/FreqPolicy/FlowPolicy is asymmetric: baselines are not conditioned on a complete reference demonstration, while DAMI receives one at both training and inference. The MAML ablation (Table VIII) shows that adding MAML improves the baselines, but those baselines still lack the reference-demo conditioning. Thus the reported gains on novel tasks could be entirely attributable to the extra input modality (an expert trajectory of the exact target task) rather than to the meta-learned initialization or the claimed 'shared skill space.' An ablation removing VMT/U2T conditioning from DAMI (or adding the same reference-demo conditioning to a baseline) is needed to separate these factors.
minor comments (4)
  1. [Fig. 2 caption; page 2] There is an untranslated Chinese paragraph directly below the Figure 2 caption in the full text. This should be removed or translated.
  2. [Table III, Table IV] Table III lists ten task columns under both 'ML10 Tasks' and 'ML45 Tasks', but Sec. IV-A states ML10 has 5 held-out tasks and ML45 has 5 held-out tasks. Clarify which tasks belong to which split and why 10 columns are shown. Table IV lacks per-task breakdown and should be completed.
  3. [Sec. III-A and Sec. III-F] The limitation about episode-level separation appears only in Sec. III-F. It should also be stated in the problem formulation (Sec. III-A) so that the reader is not misled about the training protocol when first reading the method.
  4. [Sec. IV-D] Real-world results are based on 10 trials per task; no trial-level results or confidence intervals are given. The variability for tasks like 'Object Place' (DAMI 50%, DP3 10%) is large relative to the number of trials. Report per-trial outcomes or at least a confidence interval.

Circularity Check

2 steps flagged · score 6.0 of 10

Unseen-task evaluation leaks target task through same-task reference demo; reported gains may reflect demo retrieval, not transferable dynamics.

  1. self definitional [Section III-A (Problem Formulation); Algorithm 2 lines 7 and 19]
    "A complete 200-frame reference demonstration τi is obtained by randomly selecting one complete episode from the same task dataset. ... Sample a complete reference trajectory d∼D and a batch B of observation-action windows from D."

    The framework's headline claim is generalization to 'unseen' tasks, but the evaluation protocol defines an unseen task as one for which a complete expert demonstration of that exact task is provided as a policy input. Since VMT encodes the demo's visual states and actions and U2T injects those tokens into the denoiser, the policy can succeed by retrieving/copying the demonstrated action sequence rather than by inferring transferable task dynamics. The 'unseen' success thus reduces by construction to 'imitate the supplied reference demonstration.'

  2. fitted input called prediction [Section III-F (Training and Adaptation Procedures); Section III-A]
    "The code does not enforce episode-level separation, so the selected demonstration episode may also contribute support or query windows. ... For a novel task, the same D supplies the fine-tuning windows, full reference demonstrations, and normalization statistics."

    The reference demonstration used at inference may be the very episode whose windows appear in the support/query batches during meta-training; Algorithm 2 repeats this same-dataset overlap for novel-task adaptation. The model can memorize the demonstration episode and then 'predict' its actions when that episode is presented as the reference. The reported performance on unseen tasks is therefore partly forced by the training/evaluation data construction rather than by meta-learned generalization.

full rationale

The architecture itself—Eqs. (1)–(8) and Algorithms 1–2—is a standard MAML-plus-diffusion construction whose equations are internally consistent; there is no mathematical derivation being reduced to its own output, and no load-bearing self-citation is present (all cited baselines and priors are external). The circularity lies in the empirical claim. The paper's own protocol gives the target task's complete expert demonstration as an explicit conditioning input and admits that no episode-level separation is enforced, so the reference demo can overlap with training windows. Because VMT/U2T feed the demo tokens into the action-prediction denoiser, the reported 'unseen task' success can be substantially explained by demonstration retrieval rather than by transferable dynamics. The ML45 result—where DAMI only ties DP3 (37.53%)—is consistent with this confound. This is partial circularity in the evaluation claim, not a full definitional collapse, hence score 6.

Assumptions & free parameters 5 free parameters · 5 assumptions · 3 invented entities

The paper contributes an empirical architecture; its central claim rests on domain assumptions about demonstration sufficiency, observation modality, and the transferability of MAML initialization, plus hand-chosen hyperparameters. No mathematical derivation or external theory grounds the performance.

free parameters (5)
  • Policy network weights (visual encoder, VMT, U2T, TCFM, U-Net) = Learned from demonstration data; no closed-form value
    All reported performance is produced by these fitted weights; there is no derived constant.
  • Inner-loop learning rate α = 1e-4
    Reported in Table I; hand-chosen, controls MAML adaptation.
  • Inner-loop steps = 1
    A single gradient step is used for meta-training; central to the MAML objective.
  • Support/query split sizes = 20 / 88
    Hand-chosen; defines the few-shot adaptation signal.
  • Action prediction horizon / steps = H=16, pred step 8, obs step 2
    Task-specific design choices affecting action generation.
assumptions (5)
  • domain assumption Ten complete expert demonstrations per task are sufficient to define a task and to meta-train adaptation.
    Section III-A: each D_i contains ten complete expert episodes; no evidence this number is sufficient beyond reported results.
  • domain assumption A single random complete reference demonstration from the target task is sufficient to identify task intent.
    Section III-C and Algorithm 1: τ_i is randomly selected from the same task dataset; the whole method rests on this.
  • domain assumption Point-cloud + proprioceptive state is a sufficient observation for the manipulation tasks.
    Section III-A: the policy observes a 3D point cloud and robot state only; no other sensing is used.
  • domain assumption The meta-training task distribution is related enough to novel tasks that a shared initialization transfers.
    The MAML objective relies on this transfer; the paper only tests a limited set of Meta-World/RLBench tasks.
  • domain assumption Full-MAML second-order gradients through a diffusion U-Net are stable and useful.
    Section III-B: the computation graph is retained; stability is empirical and not proven.
invented entities (3)
  • Shared skill space / task latent space
    purpose: Claims to capture task dynamics for cross-task transfer
    Internal representation with no outside falsifiable handle; only ablations support it.
  • Visual-Motor Trajectory (VMT) tokens
    purpose: Jointly encode demo visual states and actions
    Effectiveness appears only in internal component ablation (Table IX).
  • U2T fusion condition / TCFM modulation
    purpose: Fuse text, demonstration, and observation; modulate U-Net features
    No independent measurement; internal ablations only.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Dynamics-Aware Meta-Imitation for Generalization to Unseen Robotic Manipulation." pith.science (2026). https://pith.science/paper/PNXAXU6S

@misc{pith2026260715880,
  author       = {Pith},
  title        = {Pith review of: Dynamics-Aware Meta-Imitation for Generalization to Unseen Robotic Manipulation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PNXAXU6S}},
  note         = {Machine review of arXiv:2607.15880}
}
read the original abstract

Imitation Learning aims to learn skills from extensive observations and demonstrations for robots, so it suffers from data scarcity and environment generalization. The existing methods predominantly focus on imitation from in-domain tasks and consequently struggle with generalization to unseen tasks. To bridge this generalization gap, we propose the \textbf{D}ynamics-\textbf{A}ware \textbf{M}eta-\textbf{I}mitation (DAMI) framework. By integrating meta-learning to construct a shared skill space, DAMI equips agents for rapid adaptation to novel tasks. We introduce the Visual-Motor Trajectory (VMT) module to capture complex spatio-temporal dynamics within the task latent space. Furthermore, we propose the Unpaired Unified Task (U2T) block to fuse unstructured multimodal observations. To coordinate these representations, we integrate a Task-Conditioned Feature Modulation (TCFM) mechanism customized for modulating low-level 3D features. By capturing intrinsic dynamics from a random complete reference demonstration, our framework learns the underlying task logic rather than memorizing static cues, ensuring effective generalization. Extensive experiments in both simulation and real-world settings demonstrate that our approach outperforms state-of-the-art baselines regarding direct inference on seen tasks and adaptation to unseen tasks via few-shot fine-tuning.

Figures

Figures reproduced from arXiv: 2607.15880 by the authors.

Figure 1
Figure 1. Generalization gap in conventional methods. Standard approaches overfit seen data and fail on unseen tasks. The bottom sequence shows Action Misinterpretation: the robot fails to push the cylinder with an open gripper. I. INTRODUCTION Robotic manipulation aims to enable agents to interact with and change their environments [1], [2]. Imitation Learning (IL) is effective for acquiring complex robotic manipulation skil… view at source ↗
Figure 2
Figure 2. Overview of the DAMI Architecture. Given the current observation, a language instruction, and a complete reference demonstration, DAMI predicts an action trajectory using a meta-learned 3D diffusion policy. environments [20]–[22]. Within this area, diffusion policies have emerged as a powerful paradigm for modeling multimodal action distributions and generating complex visuomotor behav￾iors [23]–[25]. However, their… view at source ↗
Figure 3
Figure 3. The illustration of VMT. perception and execution, these methods are still mainly task￾specific or in-domain, with limited support for data-efficient adaptation to unseen tasks. C. Multi-task Robotic Manipulation Multi-task robotic manipulation requires effective geometric and semantic features from high-dimensional observations. RVT-2 [45] improves precision with a multi-view transformer, and GNFactor [46] uses neu… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison on Meta-World. Execution frames of DP3 (top) and DAMI (bottom) on base tasks (left) and novel-task adaptation (right). data quality or quantity, ensuring that performance differences are solely attributable to the policy architectures. Baselines.…
Figure 5
Figure 5. Figure 5: Real-world setup and tasks. The left panel shows the UR10e manipulator with an AG95 gripper, while the right panels show representative execution frames for training and unseen tasks. TABLE VI DETAILED SUCCESS RATES (%) ON ALL 45 BASE SIMULATION TASKS FROM THE ML45 BEN…
Figure 6
Figure 6. Figure 6: Initial environmental configurations for the seven real-world manipulation tasks used in the evaluation. The figure illustrates the starting states for the [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: , DAMI maintains a 100% success rate across the tested object configurations, whereas DP3 fails in the majority of trials. The qualitative results further reveal an Action Misinterpretation failure in DP3: it incorrectly closes the gripper and applies downward force, e…
Figure 8
Figure 8. Figure 8: Generalization on Push Cube. The top row compares the purple cube used for fine-tuning with the unseen Rubik’s cube used for evaluation. The bottom row shows DP3’s action misinterpretation during evaluation. “Test config” denotes trials per object. In the 200 update st…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 4 linked inside Pith

  1. [1]

    Robobrain: A unified brain model for robotic manipulation from abstract to concrete,

    Y . Ji, H. Tan, J. Shi, X. Hao, Y . Zhang, H. Zhang, P. Wang, M. Zhao, Y . Mu, P. An,et al., “Robobrain: A unified brain model for robotic manipulation from abstract to concrete,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 1724–1734

    2025

  2. [2]

    Robotic manipulation via imitation learning: Taxonomy, evolution, benchmark, and challenges,

    Z. Li, A. Chapin, E. Xiang, R. Yang, B. Machado, N. Lei, E. Dellandrea, D. Huang, and L. Chen, “Robotic manipulation via imitation learning: Taxonomy, evolution, benchmark, and challenges,” arXiv preprint arXiv:2508.17449, 2025. [Online]. Available: https: //arxiv.org/abs/2508.17449

    arXiv 2025

  3. [3]

    Adaptive video-conditioned imitation learning via bidirectional cross-domain skill transfer,

    Z. Lin, Y . Chen, Z. Li, X. Zhang, B. Liang, and Z. Liu, “Adaptive video-conditioned imitation learning via bidirectional cross-domain skill transfer,”IEEE Transactions on Automation Science and Engineering, vol. 22, pp. 23 214–23 227, 2025

    2025

  4. [4]

    Bikc+: Bimanual hierarchi- cal imitation with keypose-conditioned coordination-aware consistency policies,

    H. Xu, Y . Chen, D. Yu, Y . Ren, and J. Pan, “Bikc+: Bimanual hierarchi- cal imitation with keypose-conditioned coordination-aware consistency policies,”IEEE Transactions on Automation Science and Engineering, vol. 23, pp. 1064–1079, 2025

    2025

  5. [5]

    Toward reliable imitation learning with limited expert demonstrations via search-based inverse dynamic learning,

    Z. Lin, Z. Chen, G. Zhu, L. Wang, and J. Li, “Toward reliable imitation learning with limited expert demonstrations via search-based inverse dynamic learning,”IEEE Transactions on Automation Science and Engineering, 2026

    2026

  6. [6]

    Robot deformable object manipulation via nmpc-generated demonstrations in deep reinforcement learning,

    H. Wang, Z. Dong, T. Zhu, H. Lei, W. Shi, Z. Zhang, W. Luo, W. Wan, X. Chen, and J. Huang, “Robot deformable object manipulation via nmpc-generated demonstrations in deep reinforcement learning,”IEEE Transactions on Automation Science and Engineering, vol. 22, pp. 23 566– 23 578, 2025

    2025

  7. [7]

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

    Z. Jiang, Y . Xie, K. Lin, Z. Xu, W. Wan, A. Mandlekar, L. J. Fan, and Y . Zhu, “Dexmimicgen: Automated data generation for bimanual dexterous manipulation via imitation learning,” in2025 IEEE Interna- tional Conference on Robotics and Automation (ICRA). IEEE, 2025, pp. 16 923–16 930

    2025

  8. [8]

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

    S. Xia, H. Fang, C. Lu, and H.-S. Fang, “Cage: Causal attention enables data-efficient generalizable robotic manipulation,” in2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2025, pp. 13 242–13 249

    2025

Show all 57 references
  1. [9]

    Freqpolicy: Efficient flow-based visuomotor policy via frequency consistency,

    Y . Su, N. Liu, D. Chen, Z. Zhao, K. Wu, M. Li, Z. Xu, Z. Che, and J. Tang, “Freqpolicy: Efficient flow-based visuomotor policy via frequency consistency,”Advances in Neural Information Processing Systems, vol. 38, pp. 27 769–27 797, 2026. 12

    2026

  2. [10]

    Rvt: Robotic view transformer for 3d object manipulation,

    A. Goyal, J. Xu, Y . Guo, V . Blukis, Y .-W. Chao, and D. Fox, “Rvt: Robotic view transformer for 3d object manipulation,” inConference on Robot Learning. PMLR, 2023, pp. 694–710

    2023

  3. [11]

    Act3d: 3d feature field transformers for multi-task robotic manipulation,

    T. Gervet, Z. Xian, N. Gkanatsios, and K. Fragkiadaki, “Act3d: 3d feature field transformers for multi-task robotic manipulation,” inProceedings of The 7th Conference on Robot Learning. PMLR, 2023, pp. 3949–3965

    2023

  4. [12]

    Octo: An open-source generalist robot policy,

    Octo Model Team, D. Ghosh, H. Walke, K. Pertsch, K. Black, O. Mees, S. Dasari, J. Hejna, T. Kreiman, C. Xu, J. Luo, Y . Tan, L. Chen, P. R. Sanketi, Q. Vuong, T. Xiao, D. Sadigh, C. Finn, and S. Levine, “Octo: An open-source generalist robot policy,” inProceedings of Robotics:...

    2024

  5. [13]

    Vividex: Learning vision-based dexterous manipulation from human videos,

    Z. Chen, S. Chen, E. Arlaud, I. Laptev, and C. Schmid, “Vividex: Learning vision-based dexterous manipulation from human videos,” in2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2025, pp. 3336–3343

    2025

  6. [14]

    Dreamgen: Unlocking generalization in robot learning through video world models,

    J. Jang, S. Ye, Z. Lin, J. Xiang, J. Bjorck, Y . Fang, F. Hu, S. Huang, K. Kundalia, Y .-C. Lin,et al., “Dreamgen: Unlocking generalization in robot learning through video world models,” inProceedings of The 9th Conference on Robot Learning. PMLR, 2025, pp. 5170–5194

    2025

  7. [15]

    Airexo-2: Scaling up generalizable robotic imitation learning with low-cost exoskeletons,

    H. Fang, C. Wang, Y . Wang, J. Chen, S. Xia, J. Lv, Z. He, X. Yi, Y . Guo, X. Zhan,et al., “Airexo-2: Scaling up generalizable robotic imitation learning with low-cost exoskeletons,” inProceedings of The 9th Conference on Robot Learning. PMLR, 2025, pp. 198–220

    2025

  8. [16]

    One-shot manipulation strategy learning by making contact analogies,

    Y . Liu, J. Mao, J. B. Tenenbaum, T. Lozano-P ´erez, and L. P. Kaelbling, “One-shot manipulation strategy learning by making contact analogies,” in2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2025, pp. 15 387–15 393

    2025

  9. [17]

    One-shot video imitation via parameterized symbolic abstraction graphs,

    J. Wang, K. Liu, D. Guo, Z. Xian, and C. G. Atkeson, “One-shot video imitation via parameterized symbolic abstraction graphs,” in2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2025, pp. 10 552–10 560

    2025

  10. [18]

    Robust instant policy: Leveraging student’s t-regression model for robust in-context imitation learning of robot manipulation,

    H. Oh, A. M. Salcedo-V ´azquez, I. G. Ramirez-Alpizar, and Y . Domae, “Robust instant policy: Leveraging student’s t-regression model for robust in-context imitation learning of robot manipulation,” in2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IR...

    2025

  11. [19]

    Instant policy: In-context imitation learning via graph diffusion,

    V . V osylius and E. Johns, “Instant policy: In-context imitation learning via graph diffusion,” inThe Thirteenth International Conference on Learning Representations (ICLR), 2025

    2025

  12. [20]

    Robot dexterous grasping in cluttered scenes based on single-view point cloud,

    Q. Zhao, M. Zheng, Z. Li, S. Huang, and W. Shi, “Robot dexterous grasping in cluttered scenes based on single-view point cloud,”IEEE Transactions on Automation Science and Engineering, 2025

    2025

  13. [21]

    Region-aware grasping for stacked workpieces: A 6d-wise label self- generation method and robust evaluation strategy,

    X. Zhong, T. Gong, J. Yu, J. Luo, C. Zhou, X. Zhong, and Q. Liu, “Region-aware grasping for stacked workpieces: A 6d-wise label self- generation method and robust evaluation strategy,”IEEE Transactions on Automation Science and Engineering, vol. 22, pp. 16 899–16 912, 2025

    2025

  14. [22]

    Fast and accurate category-level object pose estimation without shape priors for robotic grasp detection,

    S. Yu, J. Yin, D.-H. Zhai, and Y . Xia, “Fast and accurate category-level object pose estimation without shape priors for robotic grasp detection,” IEEE Transactions on Automation Science and Engineering, 2025

    2025

  15. [23]

    Diffusion policy: Visuomotor policy learning via action diffusion,

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

    2025

  16. [24]

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

    Y . Ze, G. Zhang, K. Zhang, C. Hu, M. Wang, and H. Xu, “3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations,” inProceedings of Robotics: Science and Systems (RSS), 2024

    2024

  17. [25]

    Generalizable humanoid manipulation with 3d diffusion policies,

    Y . Ze, Z. Chen, W. Wang, T. Chen, X. He, Y . Yuan, X. B. Peng, and J. Wu, “Generalizable humanoid manipulation with 3d diffusion policies,” in2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2025, pp. 2873–2880

    2025

  18. [26]

    Adp: Adaptive diffusion policy energizes robots thinking in both learning and practice,

    D. Chen, Z. Chen, X. Zheng, W. Xu, C. Ma, and C. Mao, “Adp: Adaptive diffusion policy energizes robots thinking in both learning and practice,” IEEE Transactions on Automation Science and Engineering, 2025

    2025

  19. [27]

    Robot few-shot manipulation skills learning based on meta imitation learning and mixture of experts model,

    J. Zhao, J. Liu, X. Su, B. He, and S. Jiang, “Robot few-shot manipulation skills learning based on meta imitation learning and mixture of experts model,”IEEE Transactions on Automation Science and Engineering, 2026

    2026

  20. [28]

    Model-agnostic meta-learning for fast adaptation of deep networks,

    C. Finn, P. Abbeel, and S. Levine, “Model-agnostic meta-learning for fast adaptation of deep networks,” inInternational conference on machine learning. PMLR, 2017, pp. 1126–1135

    2017

  21. [29]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” inInternational Conference on Medical image computing and computer-assisted intervention. Springer, 2015, pp. 234–241

    2015

  22. [30]

    Film: Visual reasoning with a general conditioning layer,

    E. Perez, F. Strub, H. De Vries, V . Dumoulin, and A. Courville, “Film: Visual reasoning with a general conditioning layer,” inProceedings of the AAAI conference on artificial intelligence, 2018

    2018

  23. [31]

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

    E. Jang, A. Irpan, M. Khansari, D. Kappler, F. Ebert, C. Lynch, S. Levine, and C. Finn, “Bc-z: Zero-shot task generalization with robotic imitation learning,” inConference on Robot Learning. PMLR, 2022, pp. 991– 1002

    2022

  24. [32]

    Ram: Retrieval-based affordance transfer for generalizable zero-shot robotic manipulation,

    Y . Kuang, J. Ye, H. Geng, J. Mao, C. Deng, L. Guibas, H. Wang, and Y . Wang, “Ram: Retrieval-based affordance transfer for generalizable zero-shot robotic manipulation,”arXiv preprint arXiv:2407.04689, 2024

    2024 arXiv

  25. [33]

    Omnimanip: Towards general robotic manipulation via object-centric interaction primitives as spatial constraints,

    M. Pan, J. Zhang, T. Wu, Y . Zhao, W. Gao, and H. Dong, “Omnimanip: Towards general robotic manipulation via object-centric interaction primitives as spatial constraints,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 17 359–17 369

    2025

  26. [34]

    FUNCanon: Learning pose-aware action primitives via functional object canonicalization for generalizable robotic manipulation,

    H. Xu, L. Zhang, X. Hu, B. Zhong, K. Bai, Z.-C. M ´arton, Z. Bing, Z. Chen, A. C. Knoll, and J. Zhang, “FUNCanon: Learning pose-aware action primitives via functional object canonicalization for generalizable robotic manipulation,”arXiv preprint arXiv:2509.19102, 2025. [Online...

    2025 arXiv

  27. [35]

    One-shot imitation under mismatched execution,

    K. Kedia, P. Dan, A. Chao, M. A. Pace, and S. Choudhury, “One-shot imitation under mismatched execution,” in2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2025, pp. 15 649–15 656

    2025

  28. [36]

    Cova- il: Zero-shot imitation learning via contrastive viewpoint alignment on object-centric representation,

    J. Luo, S. Yang, C. Zheng, J. Fan, Y . Gao, R. Song, and W. Zhang, “Cova- il: Zero-shot imitation learning via contrastive viewpoint alignment on object-centric representation,”IEEE Transactions on Automation Science and Engineering, 2026

    2026

  29. [37]

    Autolfd: Closing the loop for learning from demonstrations,

    S. Wu, Y . Wang, and Y . Huang, “Autolfd: Closing the loop for learning from demonstrations,”IEEE Transactions on Automation Science and Engineering, vol. 22, pp. 11 124–11 138, 2025

    2025

  30. [38]

    Pointnet++: Deep hierarchical feature learning on point sets in a metric space,

    C. R. Qi, L. Yi, H. Su, and L. J. Guibas, “Pointnet++: Deep hierarchical feature learning on point sets in a metric space,”Advances in neural information processing systems, vol. 30, 2017

    2017

  31. [39]

    4d visual pre-training for robot learning,

    C. Hou, Y . Ze, Y . Fu, Z. Gao, S. Hu, Y . Yu, S. Zhang, and H. Xu, “4d visual pre-training for robot learning,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 8451–8461

    2025

  32. [40]

    Spatial- temporal aware visuomotor diffusion policy learning,

    Z. Liu, Y . Wang, K. Wang, L. Liang, X. Xue, and Y . Fu, “Spatial- temporal aware visuomotor diffusion policy learning,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 7122–7131

    2025

  33. [41]

    Rasp: Robot active scene perception with joint viewpoint planning and depth completion in cluttered environments,

    Y . Liu, T. Jia, H. Zhang, G. Yang, H. Wang, and D. Chen, “Rasp: Robot active scene perception with joint viewpoint planning and depth completion in cluttered environments,”IEEE Transactions on Automation Science and Engineering, 2025

    2025

  34. [42]

    Dual graph attention networks for multi-view visual manipulation relationship detection and robotic grasping,

    M. Ding, Y . Liu, Y . Shi, X. Lan, and N. Zheng, “Dual graph attention networks for multi-view visual manipulation relationship detection and robotic grasping,”IEEE Transactions on Automation Science and Engineering, 2025

    2025

  35. [43]

    Learning 6-dof fine-grained grasp detection based on part affordance grounding,

    Y . Song, P. Sun, P. Jin, Y . Ren, Y . Zheng, Z. Li, X. Chu, Y . Zhang, T. Li, and J. Gu, “Learning 6-dof fine-grained grasp detection based on part affordance grounding,”IEEE Transactions on Automation Science and Engineering, 2025

    2025

  36. [44]

    Affordance-guided robotic grasping via multimodal large language model reasoning,

    Z. Zhao, J. Gao, and D. Zheng, “Affordance-guided robotic grasping via multimodal large language model reasoning,”IEEE Transactions on Automation Science and Engineering, 2026

    2026

  37. [45]

    Rvt-2: Learning precise manipulation from few demonstrations,

    A. Goyal, V . Blukis, J. Xu, Y . Guo, Y .-W. Chao, and D. Fox, “Rvt-2: Learning precise manipulation from few demonstrations,” inProceedings of Robotics: Science and Systems, 2024

    2024

  38. [46]

    Gnfactor: Multi-task real robot learning with generalizable neural feature fields,

    Y . Ze, G. Yan, Y .-H. Wu, A. Macaluso, Y . Ge, J. Ye, N. Hansen, L. E. Li, and X. Wang, “Gnfactor: Multi-task real robot learning with generalizable neural feature fields,” inConference on robot learning. PMLR, 2023, pp. 284–301

    2023

  39. [47]

    Manigaussian: Dynamic gaussian splatting for multi-task robotic manipulation,

    G. Lu, S. Zhang, Z. Wang, C. Liu, J. Lu, and Y . Tang, “Manigaussian: Dynamic gaussian splatting for multi-task robotic manipulation,” in European Conference on Computer Vision. Springer, 2024, pp. 349– 366

    2024

  40. [48]

    3d gaussian splatting for real-time radiance field rendering

    B. Kerbl, G. Kopanas, T. Leimk ¨uhler, and G. Drettakis, “3d gaussian splatting for real-time radiance field rendering.”ACM Trans. Graph., vol. 42, no. 4, pp. 139–1, 2023

    2023

  41. [49]

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

    T.-W. Ke, N. Gkanatsios, and K. Fragkiadaki, “3d diffuser actor: Policy diffusion with 3d scene representations,”arXiv preprint arXiv:2402.10885, 2024

    2024 arXiv

  42. [50]

    Pdfactor: Learning tri-perspective view policy diffusion field for multi- task robotic manipulation,

    J. Tian, L. Wang, S. Zhou, S. Wang, J. Li, H. Sun, and W. Tang, “Pdfactor: Learning tri-perspective view policy diffusion field for multi- task robotic manipulation,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 15 757–15 767. 13

    2025

  43. [51]

    Flowram: Grounding flow matching policy with region-aware mamba framework for robotic manipulation,

    S. Wang, L. Wang, S. Zhou, J. Tian, J. Li, H. Sun, and W. Tang, “Flowram: Grounding flow matching policy with region-aware mamba framework for robotic manipulation,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 12 176–12 186

    2025

  44. [52]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark,et al., “Learning transferable visual models from natural language supervision,” inInternational conference on machine learning. PmLR, 2021, pp. 8748–8763

    2021

  45. [53]

    Meta-world: A benchmark and evaluation for multi-task and meta reinforcement learning,

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

    2020

  46. [54]

    RLBench: The robot learning benchmark & learning environment,

    S. James, Z. Ma, D. R. Arrojo, and A. J. Davison, “RLBench: The robot learning benchmark & learning environment,”IEEE Robotics and Automation Letters, vol. 5, no. 2, pp. 3019–3026, 2020

    2020

  47. [55]

    Flowpolicy: Enabling fast and robust 3d flow-based policy via consistency flow matching for robot manipulation,

    Q. Zhang, Z. Liu, H. Fan, G. Liu, B. Zeng, and S. Liu, “Flowpolicy: Enabling fast and robust 3d flow-based policy via consistency flow matching for robot manipulation,” inProceedings of the AAAI Conference on Artificial Intelligence, 2025, pp. 14 754–14 762

    2025

  48. [56]

    Freqpolicy: Frequency autoregressive visuomo- tor policy with continuous tokens,

    Y . Zhong, Y . Liu, C. Xiao, Z. Yang, Y . Wang, Y . Zhu, Y . Shi, Y . Sun, X. Zhu, and Y . Ma, “Freqpolicy: Frequency autoregressive visuomo- tor policy with continuous tokens,”Advances in Neural Information Processing Systems, vol. 38, pp. 56 493–56 526, 2026

    2026

  49. [57]

    Mamba policy: Towards efficient 3d diffusion policy with hybrid selective state models,

    J. Cao, Q. Zhang, J. Sun, J. Wang, H. Cheng, Y . Li, J. Ma, K. Wu, Z. Xu, Y . Shao,et al., “Mamba policy: Towards efficient 3d diffusion policy with hybrid selective state models,” in2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2025, p...

    2025

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.