Pith. sign in

REVIEW 3 major objections 6 minor 50 references

Time-Unified Diffusion Policy with Action Discrimination for Robotic Manipulation

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper argues that a diffusion policy for robotic manipulation can replace timestep-dependent denoising with a single time-unified velocity field, and that adding an action-discrimination branch makes generation both faster and more…

desk verdict The sign error in Eq. (11) makes the training target double inside the label neighborhood, so the algorithm as written reflects instead of converging; the paper needs correction but the core idea is worth reviewing. read the letter →

arxiv 2506.09422 v1 pith:3K5RTM6K submitted 2025-06-11 cs.RO cs.LG

Ye Niu , Sanping Zhou , Yizhe Li , Ye Den , Le Wang This is my paper
classification cs.ROcs.LG
keywords roboticmanipulationdiffusionpolicytime-unifiedvelocityfieldactiondiscriminationimitationlearningdenoisingRLBenchreal-timegeneration
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

Diffusion-based robot policies typically spend many iterations denoising an action, and each iteration requires the network to know which timestep it is in; this paper claims that whole schedule is unnecessary. The Time-Unified Diffusion Policy (TUDP) builds a single velocity field in action space that points any noisy action toward the nearest successful demonstration, using an action-discrimination branch to suppress interference from other successful actions. On RLBench the method reports state-of-the-art average success rates of 82.6% with multiple camera views and 83.8% with a single view, and it degrades far less than the 3D Diffuser Actor when the number of denoising iterations is cut from 100 to 5 or 10, which brings action generation close to real time. The paper also reports successful execution on six real-world tabletop tasks. If the claim holds, diffusion policies become practical for closed-loop manipulation without giving up their ability to represent multiple valid actions.

What carries the argument

The central object is the time-unified velocity field $\epsilon_x(y)$ in end-effector action space, built from conditional fields that point toward each labelled successful action $\hat{y}_i$ and weighted by a correlation factor that is zero when $y$ lies within radius $l$ of any other successful action. A companion action-discrimination network $s_\theta(x,y)$ is trained in the first stage to output the neighborhood score, and in the second stage it supplies the weight through $\lambda(y,\hat{y}) = 1 - s_\theta(x,y)\,\mathrm{sgn}(\|\hat{y}-y\|-l)$ while the unified diffusion network learns the field with an action-weighted noise loss. During inference, denoising is the fixed-point update $y_{t+1} = y_t - \epsilon_\vartheta(x,y_t)$ with early termination, and no timestep is fed to the network.

What would settle it

Measure the overlap of successful-action neighborhoods on a task with many valid poses: for the chosen $l=0.1$, compute the fraction of demonstration pairs with $\|\hat{y}_i-\hat{y}_j\| < 2l$. If that fraction is substantial, or if TUDP's success rate on such a task falls below the 3D Diffuser Actor's while reducing $l$ to eliminate overlap does not restore the reported advantage, the sparsity assumption behind the correlation weight is the point of failure.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the time indexing of a diffusion policy is not a necessary scaffold but a source of confusion. TUDP defines a single time-unified velocity field $\epsilon_x(y)$ whose value at any point in action space does not depend on how many denoising steps have elapsed. The field is a mixture of conditional fields pointing to the labeled successful actions, with the correlation weight $\lambda(y,\hat{y}_i)$ zeroing out the influence of every successful action whose neighborhood is not the one the noisy action belongs to. An action-discrimination network learns that membership from the scene observation, and a two-stage training procedure first fits the discriminator and then fits the unified diffusion network against an action-weighted noise loss. With this field, iterative denoising becomes $y_{t+1} = y_t - \epsilon_\vartheta(x, y_t)$ plus early stopping, and the same network can run for 100 iterations or 5. The paper claims this both shortens training and inference and improves action accuracy, and it supports the claim with state-of-the-art RLBench numbers, few-iteration comparisons against the 3D Diffuser Actor, ablations, and real-robot tests.

Load-bearing premise

The load-bearing premise is that successful actions for a given scene are sparse enough that radius-$l$ neighborhoods around them do not overlap; if valid actions are dense or continuous, the action-discrimination weight cannot separate them and the unified velocity field degenerates.

Editorial extensions

If this is right

  • At low iteration counts (5, 10, 25) TUDP's success rate stays markedly above the 3D Diffuser Actor's, so the time-unified field is the mechanism that makes short-horizon action generation reliable.
  • The ablation row without action-wise training drops average success from 82.6% to 79.1%, so the action-discrimination information, not just the unified field, is carrying part of the accuracy gain.
  • TUDP's single-view average of 83.8% on 10 RLBench tasks is the strongest reported number in that setting, which would let manipulation policies run with a single front camera.
  • The real-robot experiments (51/60 successes across six tasks trained from 58 demonstrations) suggest the method transfers outside the RLBench simulator.
  • Early termination and the velocity limitation control the speed-accuracy trade-off; with all components the reported inference time is 0.43s per action, near the loop rate needed for closed-loop control.

Reading between the lines

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

  • Beyond the paper: the timestep-free field could be plugged into flow-matching or consistency-model objectives, so the speed gains of those methods and TUDP's accuracy gain would likely compound rather than compete.
  • Beyond the paper: the action score $s_\theta(x,y)$ could be reused at inference as a confidence signal, letting the robot stop, resample, or ask for help when the score stays low; the paper uses the score only during training.
  • Beyond the paper: the fixed neighborhood radius $l=0.1$ should fail on tasks with a continuous manifold of equally valid motions, such as pouring or sweeping; an action-dependent or per-task radius is the natural extension the limitation section points toward.
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 proposes Time-Unified Diffusion Policy (TUDP), a diffusion-based action generation method for robotic manipulation. The key ideas are (i) a time-unified velocity field in action space that points from noisy actions toward a successful action, with a velocity-limitation hyperparameter v, and (ii) an action-wise training scheme in which an action-discrimination network estimates whether a noisy action lies inside a successful action's neighborhood and uses this score to weight the velocity regression loss. The method is evaluated on 18 RLBench tasks in multi-view and single-view settings, compared with several baselines, ablated for its components and hyperparameters, and demonstrated on six real-world tasks. The authors report state-of-the-art average success rates of 82.6% (multi-view) and 83.8% (single-view), and improved success at low denoising iterations.

Significance. If the method performs as reported, the paper offers a practical speed-up for diffusion-based policies in robotic manipulation, with strong benchmark results and real-robot validation. The iteration-count comparison in Figure 6 directly tests the efficiency claim, and the ablations cover the main components. However, the central derivation contains a sign/definition issue in Eq. (11) that makes the described training target inconsistent with the intended velocity field; until this is resolved, the reported empirical results are not explained by the published algorithm. The paper would be a useful contribution after the derivation is corrected and the inference-time initialization is clarified.

major comments (3)
  1. [3.3, Eqs. (11) and (13)] As written, the correlation weight approximation is inconsistent with Eq. (5) and makes the training target wrong. For a noisy action y inside the label action's neighborhood (||y−ŷ||<l), Eq. (9) gives sθ(x,y)≈1 and sgn(||ŷ−y||−l)=−1, so Eq. (11) yields λ≈2. The action-weighted loss in Eq. (13) then regresses the unified diffusion network to 2ε(y|ŷ) inside the neighborhood instead of ε(y|ŷ). Consequently, Algorithm 1 with the learned field performs y_{t+1}=y_t−2(y_t−ŷ)=2ŷ−y_t for v=1, i.e., a reflection about the successful action, and the early termination condition (step size 2||y_t−ŷ||≥δ) never triggers. In a one-dimensional single-action example (ŷ=1, v=1, l=0.1), the iterates oscillate between values symmetric about 1 and do not converge, contradicting the one-step convergence claim in §3.2. Moreover, Eq. (11) does not approximate Eq. (5): for y inside another successful action's neighborhood, sθ≈0 and sgn(||ŷ−y||−l)=+1, giving λ≈1, whereas Eq. (5) requires λ=0. Please correct the sign/definition in Eq. (11) or provide the implementation detail that resolves this discrepancy.
  2. [4.1 and Algorithm 1] The distribution of the initial noisy action y0 at inference is not specified. Training noisy actions are drawn from p(y|ŷ)=N(ŷ,σ^2) with σ=0.5 (Eq. (8) and Table 5), so the network has essentially no training signal more than a few standard deviations away from the demonstrations. If y0 is sampled from a standard Gaussian as is common for diffusion policies, the method relies on extrapolation to regions never seen in training; the assertion that p(y|ŷ)>0 'everywhere in action space' is a formal property of the Gaussian, not a guarantee that the learned field is meaningful there. Please state the initialization distribution and, if it is standard normal, either match the training noise distribution to the inference range (e.g., by sampling y from a wider prior) or provide quantitative evidence (e.g., velocity-field error as a function of distance from demonstrations) that the network generalizes as required.
  3. [3.2, after Eq. (7)] The statement that 'the neighborhoods of successful actions can converge to the corresponding successful action in one step' is a direct property of the chosen target field ε(y|ŷ)=y−ŷ for ||y−ŷ||≤v, not an emergent property of the method. The paper should present this as a design choice and then validate that the trained network actually reproduces the analytic field in the relevant region. The qualitative velocity-field visualization in Figure 7 is supportive but does not quantify the fit; I recommend adding a quantitative evaluation of the learned field (e.g., mean absolute error between ϵϑ(x,y) and the intended target on held-out scenes) to confirm the mechanism behind the success-rate results.
minor comments (6)
  1. [Algorithm 2, line 6] The expression 'p(y|ˆt)' should read 'p(y|ŷ)'.
  2. [Table 1] The column header 'Turen Tap' should be 'Turn Tap'.
  3. [Section 5, Limitation] The text 'Time Unifie Diffusion Policy' should be 'Time-Unified Diffusion Policy'.
  4. [Section 4.3, Table 5] The table header says 'Gaussian Variance' but the values 1, 0.5, etc. are used as σ in Eq. (8), which is the standard deviation; please clarify the notation to avoid ambiguity about the actual noise scale.
  5. [Section 4.3, paragraph on Neighborhood Radius] The phrase 'a neighborhood radius for elementary schools can increase the learning difficulty' is unclear; it likely means 'a very small neighborhood radius increases the learning difficulty'.
  6. [Section 4.4] The phrase 'the dynamic radius schedule' is not defined in the method; if it refers to the early-termination threshold or velocity limitation, it should be described in Section 3.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the core training targets are constructed from demonstration labels in a standard supervised manner, and the headline claims are measured against external benchmarks rather than derived from the method's own outputs.

full rationale

TUDP's derivation chain is self-contained in the circularity sense. The time-unified velocity field is constructed as a regression target from the demonstration labels (Eqs. 2, 6, and 13), and the action-discrimination branch is an auxiliary classifier trained on the same labels (Eqs. 9 and 10). These are ordinary supervised training objectives, not predictions derived from the method's own outputs. The load-bearing empirical claims—SOTA RLBench success rates and better performance at low denoising iterations—are measured against external baselines in Tables 1 and 2 and Figure 6, and are falsifiable. Hyperparameters such as the neighborhood radius l, Gaussian variance σ, and velocity cap v are tuned on the benchmark, but this is standard model selection, not a renamed fit or an output that is forced by definition. There are no load-bearing self-citations and no imported uniqueness theorems. The paper also explicitly acknowledges a real limitation in Section 5: the neighborhood-radius assumption fails in dense interaction scenarios, and the method requires several tuned hyperparameters. Separately, the paper contains an internal sign inconsistency that is worth flagging: with Eq. (9) giving sθ≈1 inside the label neighborhood, Eq. (11) yields λ≈2, contradicting the intended λ=1 from Eq. (5); if implemented literally, Algorithm 1 would reflect rather than converge in a one-dimensional example. That is a correctness or reproducibility flaw, not a circularity, because it does not make any claimed result equivalent to its inputs. No circular step meets the quoted-reduction bar, so the circularity score is 0.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central construction rests on several design choices that are not derived from first principles: the neighborhood radius l and Gaussian variance σ are selected by ablation, the velocity cap v is chosen to balance speed and accuracy, and the score sharpness m and opening weight w_open are fixed by hand. The method assumes successful actions are sparse, that the Gaussian training noise covers the action space, that Euclidean distance is valid for the action representation including rotations, and that the network generalizes outside its training distribution. No new physical entities are introduced; the auxiliary action-score function is a learned latent, not an independent entity.

free parameters (6)
  • Neighborhood radius l = 0.1
    Tuned via ablation over 0.05, 0.1, 0.2, 0.4 (Table 4); sets the action-score label boundary and the correlation weight in Eq. (5).
  • Gaussian variance σ = 0.5
    Tuned via ablation over 0.05, 0.15, 0.25, 0.5, 1.0 (Table 5); controls the noisy-action sampling distribution p(y|ŷ).
  • Velocity limitation v = 1
    Caps the magnitude of the conditional velocity field in Eq. (2); Figure 6 compares v=0.5 and v=1, and Table 3 shows the variant with velocity limitation is the final method.
  • Score sharpness m = m < -10
    Set to a value below -10 in Eq. (9); the exact value is not reported.
  • Opening loss weight w_open = 0.4
    Fixed weight for the gripper opening loss in Eq. (14).
  • Early termination threshold δ = 0.01
    Used in Algorithm 1 to stop denoising when the action update is small.
assumptions (5)
  • domain assumption Successful actions are sparse and their l-neighborhoods do not overlap
    Required for Eq. (5) to separate successful actions; the authors state in the Limitation that dense interaction scenarios break this assumption.
  • domain assumption Noisy actions are sampled from a Gaussian around the label action, p(y|ŷ) = N(ŷ, σ²I), and this distribution covers the action space
    Section 3.3, Eq. (8); in practice σ=0.5 concentrates training near demonstrations, so the 'greater than 0 everywhere' claim is theoretical.
  • domain assumption The action space is treated as Euclidean for noise and distance computations, including rotations
    The paper concatenates translation and rotation into y and uses Euclidean norms in Eqs. (2)-(9) without describing an SO(3) distance.
  • domain assumption The learned network approximates the time-unified field across the full action space, including out-of-distribution initial noise
    Inference starts from standard Gaussian noise (Algorithm 1), but training samples y only within a few σ of demonstrations.
  • domain assumption Pretrained CLIP encoders provide features sufficient for aligning observations and actions
    The multimodal encoders are frozen and 'training-free', cited in Section 3.1.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Time-Unified Diffusion Policy with Action Discrimination for Robotic Manipulation." pith.science (2026). https://pith.science/paper/3K5RTM6K

@misc{pith2026250609422,
  author       = {Pith},
  title        = {Pith review of: Time-Unified Diffusion Policy with Action Discrimination for Robotic Manipulation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3K5RTM6K}},
  note         = {Machine review of arXiv:2506.09422}
}
read the original abstract

In many complex scenarios, robotic manipulation relies on generative models to estimate the distribution of multiple successful actions. As the diffusion model has better training robustness than other generative models, it performs well in imitation learning through successful robot demonstrations. However, the diffusion-based policy methods typically require significant time to iteratively denoise robot actions, which hinders real-time responses in robotic manipulation. Moreover, existing diffusion policies model a time-varying action denoising process, whose temporal complexity increases the difficulty of model training and leads to suboptimal action accuracy. To generate robot actions efficiently and accurately, we present the Time-Unified Diffusion Policy (TUDP), which utilizes action recognition capabilities to build a time-unified denoising process. On the one hand, we build a time-unified velocity field in action space with additional action discrimination information. By unifying all timesteps of action denoising, our velocity field reduces the difficulty of policy learning and speeds up action generation. On the other hand, we propose an action-wise training method, which introduces an action discrimination branch to supply additional action discrimination information. Through action-wise training, the TUDP implicitly learns the ability to discern successful actions to better denoising accuracy. Our method achieves state-of-the-art performance on RLBench with the highest success rate of 82.6% on a multi-view setup and 83.8% on a single-view setup. In particular, when using fewer denoising iterations, TUDP achieves a more significant improvement in success rate. Additionally, TUDP can produce accurate actions for a wide range of real-world tasks.

Figures

Figures reproduced from arXiv: 2506.09422 by the authors.

Figure 1
Figure 1. Comparison of different denoising processes in action space. The single￾step denoising directions are indicated by red arrows, and the complete denoising paths are indicated by blue arrows. Previous diffusion policy methods adopt the time-varying velocity field, which has unclear denoising directions and high temporal complexity. In contrast, we design the time-unified velocity field for efficient and accurate robot… view at source ↗
Figure 2
Figure 2. The architecture of TUDP. The action score s is used to identify successful actions. And the noise prediction ϵ is used to correct noisy action y. main development of our approach, including the time-unified velocity field and action-wise training method. 3.1. Overview Our TUDP adopts a keyframe-based manipulation framework, which selects discrete keyframes to compose the robot demonstrations [41]. Essentially, each… view at source ↗
Figure 3
Figure 3. Comparison of two time-unified velocity fields. Red dots denote successful actions, blue dots denote denoised actions, half-red and half-blue dots indicate that the de￾noised actions equal the successful actions, and the orange line is the probability contour of the noisy action [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Action-wise training method with two stages. We omit the multimodal encoders in our method that are training-free. converge to the corresponding successful action in one step under the guidance of the time-unified velocity field. 3.3. Action-wise Training Method As sho…
Figure 5
Figure 5. Figure 5: Overview of 14 tasks. We show the image observations in these RLBench tasks. with 14 tasks illustrated in [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: Comparison of average success rates at different iterations. When using fewer denoising iterations, our TUDP achieves a significantly higher success rate than the 3D Diffuser Actor. denoising iterations to control generation time and compare the performance of our meth…
Figure 7
Figure 7. Figure 7: Visualization of the magnitude in different velocity fields. In the same scene, we compare the time-unified velocity field of TUDP with the time-varying velocity field of the existing method. 4.3. Ablation Experiments Ablation on the Components of TUDP. We have verifie…
Figure 8
Figure 8. Figure 8: Overview of 6 real-world tasks. We show the image observations in these real￾world scenarios. ries defined by human-specified waypoints. For inference, we apply the BiRRT planner integrated with MoveIt! ROS package [50] to act towards the predicted robot poses. Visual …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 38 canonical work pages

  1. [1]

    X. Jia, D. Blessing, X. Jiang, M. Reuss, A. Donat, R. Lioutikov, G. Neu- mann, Towards diverse behaviors: A benchmark for imitation learning with humandemonstrations, in: TheTwelfthInternationalConferenceonLearn- ing Representations, 2024. 21

    work page 2024

  2. [2]

    T.-W. Ke, N. Gkanatsios, K. Fragkiadaki, 3d diffuser actor: Policy diffu- sion with 3d scene representations, in: First Workshop on Vision-Language Models for Navigation and Manipulation at ICRA 2024, 2024

    work page 2024

  3. [3]

    Yan, Y.-H

    G. Yan, Y.-H. Wu, X. Wang, Dnact: Diffusion guided multi-task 3d policy learning, arXiv preprint arXiv:2403.04115 (2024)

    arXiv 2024

  4. [4]

    T. Oba, M. Walter, N. Ukita, Read: Retrieval-enhanced asymmetric dif- fusion for motion planning, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 17974–17984

    work page 2024

  5. [5]

    G. Lu, Z. Gao, T. Chen, W. Dai, Z. Wang, Y. Tang, Manicm: Real-time 3d diffusion policy via consistency model for robotic manipulation, arXiv preprint arXiv:2406.01586 (2024)

    arXiv 2024

  6. [6]

    Zhang, Z

    Q. Zhang, Z. Liu, H. Fan, G. Liu, B. Zeng, S. Liu, Flowpolicy: Enabling fast and robust 3d flow-based policy via consistency flow matching for robot manipulation, in: Proceedings of the AAAI Conference on Artificial Intel- ligence, Vol. 39, 2025, pp. 14754–14762

    work page 2025

  7. [7]

    J. Ho, A. Jain, P. Abbeel, Denoising diffusion probabilistic models, Ad- vances in neural information processing systems 33 (2020) 6840–6851

    2020

  8. [8]

    J. Song, C. Meng, S. Ermon, Denoising diffusion implicit models, in: In- ternational Conference on Learning Representations, 2021

    work page 2021

Show all 50 references
  1. [9]

    Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, B. Poole, Score-based generative modeling through stochastic differential equations, in: International Conference on Learning Representations, 2021

    2021

  2. [10]

    Dockhorn, A

    T. Dockhorn, A. Vahdat, K. Kreis, Score-based generative modeling with critically-damped langevin diffusion, in: International Conference on Learning Representations, 2022

    2022

  3. [11]

    Jolicoeur-Martineau, K

    A. Jolicoeur-Martineau, K. Li, R. Piché-Taillefer, T. Kachman, I. Mitliagkas, Gotta go fast when generating data with score-based models, arXiv preprint arXiv:2105.14080 (2021). 22

    2021 arXiv

  4. [12]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, B. Ommer, High- resolution image synthesis with latent diffusion models, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10684–10695

    2022

  5. [13]

    Vahdat, K

    A. Vahdat, K. Kreis, J. Kautz, Score-based generative modeling in latent space, Advances in neural information processing systems 34 (2021) 11287– 11302

    2021

  6. [14]

    Davtyan, S

    A. Davtyan, S. Sameni, P. Favaro, Efficient video prediction via sparsely conditioned flow matching, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 23263–23274

    2023

  7. [15]

    Bansal, E

    A. Bansal, E. Borgnia, H.-M. Chu, J. Li, H. Kazemi, F. Huang, M. Gold- blum, J. Geiping, T. Goldstein, Cold diffusion: Inverting arbitrary image transforms without noise, Advances in Neural Information Processing Sys- tems 36 (2024)

    2024

  8. [16]

    Lipman, R

    Y. Lipman, R. T. Chen, H. Ben-Hamu, M. Nickel, M. Le, Flow matching for generative modeling, arXiv preprint arXiv:2210.02747 (2022)

    2022 arXiv

  9. [17]

    S. Lin, B. Liu, J. Li, X. Yang, Common diffusion noise schedules and sample steps are flawed, in: Proceedings of the IEEE/CVF winter conference on applications of computer vision, 2024, pp. 5404–5411

    2024

  10. [18]

    Zhang, H

    P. Zhang, H. Yin, C. Li, X. Xie, Tackling the singularities at the endpoints of time intervals in diffusion models, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 6945– 6954

    2024

  11. [19]

    X. Wu, Y. Hua, C. Liang, J. Zhang, H. Wang, T. Song, H. Guan, Cgi-dm: Digital copyright authentication for diffusion models via contrasting gra- dient inversion, in: 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), IEEE Computer Society, 2024, pp. 10...

    2024

  12. [20]

    A. Zeng, P. Florence, J. Tompson, S. Welker, J. Chien, M. Attarian, T. Armstrong, I. Krasin, D. Duong, V. Sindhwani, et al., Transporter net- works: Rearranging the visual world for robotic manipulation, in: Confer- ence on Robot Learning, PMLR, 2021, pp. 726–747

    2021

  13. [21]

    M. Zare, P. M. Kebria, A. Khosravi, S. Nahavandi, A survey of imitation learning: Algorithms, recent developments, and challenges, IEEE Transac- tions on Cybernetics (2024)

    2024

  14. [22]

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

    2024

  15. [23]

    T. Z. Zhao, V. Kumar, S. Levine, C. Finn, Learning fine-grained biman- ual manipulation with low-cost hardware, arXiv preprint arXiv:2304.13705 (2023)

    2023 arXiv

  16. [24]

    Blukis, Y.-W

    A.Goyal, J.Xu, Y.Guo, V. Blukis, Y.-W. Chao, D.Fox, Rvt: Roboticview transformer for 3d object manipulation, in: Conference on Robot Learning, PMLR, 2023, pp. 694–710

    2023

  17. [25]

    J. Ho, S. Ermon, Generative adversarial imitation learning, Advances in neural information processing systems 29 (2016)

    2016

  18. [26]

    Tsurumine, T

    Y. Tsurumine, T. Matsubara, Goal-aware generative adversarial imita- tion learning from imperfect demonstration for robotic cloth manipulation, Robotics and Autonomous Systems 158 (2022) 104264

    2022

  19. [27]

    Florence, C

    P. Florence, C. Lynch, A. Zeng, O. A. Ramirez, A. Wahid, L. Downs, A. Wong, J. Lee, I. Mordatch, J. Tompson, Implicit behavioral cloning, in: Conference on robot learning, PMLR, 2022, pp. 158–168

    2022

  20. [28]

    D.-N.Ta, E.Cousineau, H.Zhao, S.Feng, Conditionalenergy-basedmodels for implicit policies: The gap between theory and practice, arXiv preprint arXiv:2207.05824 (2022). 24

    2022 arXiv

  21. [29]

    H. Ryu, J. Kim, H. An, J. Chang, J. Seo, T. Kim, Y. Kim, C. Hwang, J. Choi, R. Horowitz, Diffusion-edfs: Bi-equivariant denoising generative modeling on se (3) for visual robotic manipulation, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognitio...

    2024

  22. [30]

    Urain, N

    J. Urain, N. Funk, J. Peters, G. Chalvatzaki, Se (3)-diffusionfields: Learn- ing smooth cost functions for joint grasp and motion optimization through diffusion, in: 2023 IEEE International Conference on Robotics and Au- tomation (ICRA), IEEE, 2023, pp. 5923–5930

    2023

  23. [31]

    Pearce, T

    T. Pearce, T. Rashid, A. Kanervisto, D. Bignell, M. Sun, R. Georgescu, S. V. Macua, S. Z. Tan, I. Momennejad, K. Hofmann, S. Devlin, Imitating human behaviour with diffusion models, in: The Eleventh International Conference on Learning Representations, 2023

    2023

  24. [32]

    C. Chi, Z. Xu, S. Feng, E. Cousineau, Y. Du, B. Burchfiel, R. Tedrake, S. Song, Diffusion policy: Visuomotor policy learning via action diffusion, The International Journal of Robotics Research (2023) 02783649241273668

    2023

  25. [33]

    H. Wu, Y. Jing, C. Cheang, G. Chen, J. Xu, X. Li, M. Liu, H. Li, T. Kong, Unleashing large-scale video generative pre-training for visual robot ma- nipulation, in: The Twelfth International Conference on Learning Repre- sentations, 2024

    2024

  26. [34]

    C. Chi, S. Feng, Y. Du, Z. Xu, E. Cousineau, B. Burchfiel, S. Song, Diffusion policy: Visuomotor policy learning via action diffusion, CoRR abs/2303.04137 (2023)

    2023 arXiv

  27. [35]

    Y. Ze, G. Zhang, K. Zhang, C. Hu, M. Wang, H. Xu, 3d diffusion pol- icy: Generalizablevisuomotorpolicylearningviasimple3drepresentations, arXiv preprint arXiv:2403.03954 (2024)

    2024 arXiv

  28. [36]

    Y. Ze, G. Zhang, K. Zhang, C. Hu, M. Wang, H. Xu, 3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations, in: 25 ICRA 2024 Workshop on 3D Visual Representations for Robot Manipula- tion, 2024

    2024

  29. [37]

    Driess, I

    D. Driess, I. Schubert, P. Florence, Y. Li, M. Toussaint, Reinforcement learning with neural radiance fields, Advances in Neural Information Pro- cessing Systems 35 (2022) 16931–16945

    2022

  30. [38]

    Chisari, N

    E. Chisari, N. Heppert, M. Argus, T. Welschehold, T. Brox, A. Valada, Learning robotic manipulation policies from point clouds with conditional flow matching, in: 8th Annual Conference on Robot Learning, 2024. URLhttps://openreview.net/forum?id=vtEn8NJWlz

    2024

  31. [39]

    N. Funk, J. Urain, J. Carvalho, V. Prasad, G. Chalvatzaki, J. Peters, Ac- tionflow: Equivariant, accurate, and efficient policies with spatially sym- metric flow matching, CoRR abs/2409.04576 (2024). URLhttps://doi.org/10.48550/arXiv.2409.04576

    2024 doi arXiv

  32. [40]

    X. Ma, S. Patidar, I. Haughton, S. James, Hierarchical diffusion policy for kinematics-aware multi-task robotic manipulation, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 18081–18090

    2024

  33. [41]

    James, K

    S. James, K. Wada, T. Laidlow, A. J. Davison, Coarse-to-fine q-attention: Efficient learning for visual robotic manipulation via discretisation, in: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 13739–13748

    2022

  34. [42]

    James, Z

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

    2020

  35. [43]

    Rohmer, S

    E. Rohmer, S. P. Singh, M. Freese, V-rep: A versatile and scalable robot simulation framework, in: 2013 IEEE/RSJ international conference on in- telligent robots and systems, IEEE, 2013, pp. 1321–1326. 26

    2013

  36. [44]

    S. Chen, R. Garcia, C. Schmid, I. Laptev, Polarnet: 3d point clouds for language-guided robotic manipulation (2023)

    2023

  37. [45]

    Guhur, S

    P.-L. Guhur, S. Chen, R. G. Pinel, M. Tapaswi, I. Laptev, C. Schmid, Instruction-driven history-aware policies for robotic manipulations, in: Conference on Robot Learning, PMLR, 2023, pp. 175–187

    2023

  38. [46]

    Shridhar, L

    M. Shridhar, L. Manuelli, D. Fox, Perceiver-actor: A multi-task trans- former for robotic manipulation, in: Conference on Robot Learning, PMLR, 2023, pp. 785–799

    2023

  39. [47]

    Gervet, Z

    T. Gervet, Z. Xian, N. Gkanatsios, K. Fragkiadaki, Act3d: 3d feature field transformers for multi-task robotic manipulation, in: 7th Annual Confer- ence on Robot Learning, 2023

    2023

  40. [48]

    Goyal, V

    A. Goyal, V. Blukis, J. Xu, Y. Guo, Y.-W. Chao, D. Fox, Rvt-2: Learn- ing precise manipulation from few demonstrations, CoRR abs/2406.08545 (2024)

    2024 arXiv

  41. [49]

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

    2023

  42. [50]

    Coleman, I

    D. Coleman, I. Sucan, S. Chitta, N. Correll, Reducing the barrier to en- try of complex robotic software: a moveit! case study, arXiv preprint arXiv:1404.3785 (2014). 27

    2014 arXiv

Pith tools

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