Pith. sign in

REVIEW 3 major objections 6 minor 45 references

A one-step flow policy can beat diffusion-based policies in robot manipulation by selectively aligning generated actions with expert demonstrations.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 22:46 UTC pith:WYI4UJ5S

load-bearing objection Promising selective-alignment heuristic for one-step flow policies, but the Algorithm 2/prose mismatch and unreported δ make the central mechanism untestable as written. the 3 major comments →

arxiv 2511.08583 v2 pith:WYI4UJ5S submitted 2025-11-11 cs.RO cs.LG

SeFA-Policy: Fast and Accurate Visuomotor Policy Learning with Selective Flow Alignment

Rong Xue , Jiageng Mao , Mingtong Zhang , Yue Wang This is my paper
classification cs.RO cs.LG
keywords visuomotor policy learningrectified flowselective flow alignmentimitation learningone-step inferenceflow matchingrobot manipulationaction consistency
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper argues that the main barrier to fast flow-based visuomotor policies is not the flow model itself but the observation–action inconsistency introduced by reflow distillation. To fix this, SeFA selectively replaces each reflow-generated action with its nearest ground-truth action from expert demonstrations when the two are close enough, while leaving genuinely novel actions untouched. The result is a one-step flow policy that is both observation-aligned and multimodal, achieving 62.3% average success across 66 simulated tasks versus 38.9% for Diffusion Policy, while cutting inference latency by 98.7%. This matters because it points to a practical route to real-time imitation learning without sacrificing accuracy or diversity.

Core claim

SeFA demonstrates that reflow-induced observation–action inconsistency can be repaired by a simple selective alignment: for each generated action, find the nearest expert action conditioned on the same observation; if the Euclidean distance is below a threshold, replace the generated action with that expert action, otherwise keep it. Training the flow network on these selectively aligned couplings yields a policy that retains the straight-line one-step sampling of rectified flow while recovering the accuracy lost during reflow, in some cases exceeding the base policy.

What carries the argument

Selective Flow Alignment: a post-processing step applied to reflow-generated training couplings that replaces a generated action with the nearest ground-truth action from expert demonstrations when their distance is below a threshold δ. This maintains a deterministic coupling between noise and action while restoring consistency with the visual observation, and preserves multimodality by not replacing actions that differ from the expert distribution.

Load-bearing premise

The method assumes that replacing a reflow-generated action with the nearest expert action (when within a distance threshold δ) preserves a valid straight-line flow coupling, so the one-step policy remains accurate; δ is chosen heuristically and not reported, so the transfer to new tasks is untested.

What would settle it

Evaluate SeFA on a task with continuous precision requirements (e.g., peg insertion) with δ varied from 0 to the maximum observed nearest-neighbor distance. If success rate drops or variance increases as δ grows beyond a small value, the nearest-neighbor replacement distorts the action distribution and the method's advantage will not transfer.

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

If this is right

  • One-step flow policies become viable for real-time robot control, with per-action inference at ~16.7 ms instead of ~1.3 s for Diffusion Policy.
  • The method can be layered on any flow-matching policy (e.g., π0) without architectural changes, as SeFA only modifies the training couplings.
  • SeFA reduces performance variance across noise seeds, making policies more predictable during evaluation.
  • The ablation results suggest reflow degradation is indeed error accumulation, not a fundamental limitation of distillation.

Where Pith is reading between the lines

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

  • Because SeFA replaces actions only within a threshold, the method implicitly assumes that reflow errors are locally corrected in action space; if that holds, a per-action or per-dimension adaptive threshold could further improve performance without tuning.
  • A testable extension is to apply SeFA to other conditional flow-matching settings (e.g., image or audio generation) where observation–output consistency matters, and measure whether selective alignment reduces distortion.
  • The reported gains may come partly from the flow-matching schedule itself rather than SeFA; a cleaner ablation that trains a flow policy without reflow but with the same data would isolate the contribution.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper introduces SeFA-Policy, a visuomotor policy learning method based on rectified flow with a 'selective flow alignment' step. The authors claim that after reflow distillation, generated actions can become inconsistent with the visual observation, and that this inconsistency can be corrected by selectively replacing reflow-generated actions with nearby expert actions. The method is evaluated on 66 simulated tasks across five benchmarks and 7 real-world tasks, reporting an average success rate of 62.3% versus 38.9% for Diffusion Policy and 41.5% for AdaFlow, while reducing inference latency to 16.72 ms per action prediction (a claimed 98.7% reduction relative to Diffusion Policy). The central technical idea is that replacing a generated action with an expert action only when they are within a distance threshold preserves the straightness of the flow paths and thus enables accurate one-step inference.

Significance. If the empirical results hold, the paper makes a useful contribution to real-time visuomotor control: it demonstrates that one-step flow-based policies can exceed the accuracy of multi-step diffusion policies on a broad set of manipulation tasks. The breadth of the evaluation (66 simulated tasks plus 7 real-robot tasks) is a genuine strength, and the code release supports reproducibility. However, the core algorithmic mechanism is not precisely specified: the text and pseudocode disagree on the search space for the selective alignment, and the key threshold parameter is never reported. The theoretical justification for optimal-transport preservation is also incorrect as stated. These issues currently prevent the reader from verifying the central claim, though they appear addressable in revision.

major comments (3)
  1. [Section III-B / Algorithm 2] There is an internal inconsistency in the description of the selective alignment search. The prose says 'we methodically search the ground-truth action dataset to identify its nearest neighbor' (action-space search), whereas Algorithm 2 says 'Find the nearest ground-truth condition to O' (observation-space search) and then compares the corresponding expert action to a_reflow_0. These are different operations. If the search is in action space, the selected expert action may come from a different observation, so the method does not restore observation-action consistency and may inject off-conditioning actions. If the search is in observation space, the distance metric for observations and the role of the threshold δ need to be defined. This ambiguity is load-bearing for the paper's central claim, and the pseudocode must be aligned with the intended algorithm.
  2. [Section III-B / Algorithm 2] The alignment threshold (δ in Algorithm 2, called ε in the text) is never reported. The paper states it is selected heuristically, but no value or sensitivity analysis is given. The claim that hyperparameters were not tuned per task is therefore not verifiable, and the experiments are not reproducible without this value. Please report the threshold(s) used for all experiments and provide an ablation or sensitivity study showing how performance varies with δ. This is a required element for the central mechanism, not a peripheral detail.
  3. [Section III-A / III-B] The statement that the base-policy coupling (a_base_T, a_base_0) 'ensures optimal transport efficiency... a property that follows directly from Jensen's inequality' is not correct. A coupling generated by a learned flow matching model from random noise does not generally satisfy the optimal transport property; Jensen's inequality alone does not establish such a claim. Consequently, the assertion in Section III-B that the selectively aligned coupling 'should maintain the property of optimal transport efficiency' is unsupported: replacing a target action with a different expert action generically changes the coupling, can increase transport cost, and may introduce crossing straight paths. The empirical result in Table V (1-step Euler matches 100-step for SeFA) is encouraging, but the theoretical justification as written is wrong and should be corrected or substantially softened.
minor comments (6)
  1. [Section III-B / Algorithm 2] The threshold symbol is inconsistent: the text uses ε, Algorithm 2 uses δ. Use a single symbol throughout.
  2. [Table I caption] The caption states 'SeFA achieves 50.1% relative improvement compared to Diffusion Policy.' The numbers in the table give (62.3−38.9)/38.9 ≈ 60.2% relative improvement over Diffusion Policy; 50.1% is the relative improvement over AdaFlow Policy (41.5%). Please correct this misstatement.
  3. [Table VII / Table I] The average success rate is reported as 62.2 in Table VII but 62.3 in Table I. Reconcile the discrepancy.
  4. [Section III-A / Algorithm 1] The time variable t is defined on [0,T] in the equations but sampled from Uniform([0,1]) in Algorithm 1. Use normalized time consistently to avoid confusion.
  5. [Section IV-B / Table IV] The statement 'π0 [26] is equivalent to the first policy in Table IV' is an overclaim. π0 is a large vision-language-action model with pretraining and a different architecture; it is not equivalent to the base policy described here unless you mean only the flow-matching loss. Please clarify what 'equivalent' means.
  6. [Algorithm 2] The notation a_reflow,*_0 is not defined. Identify it explicitly as the expert action associated with the nearest ground-truth condition.

Circularity Check

0 steps flagged

No significant circularity: SeFA is a standard rectified-flow distillation pipeline with expert-action correction; its success claims are closed-loop task outcomes, not quantities re-inserted as predictions.

full rationale

SeFA's derivation chain is base-policy flow training -> reflow distillation on self-generated couplings -> selective replacement of some reflow actions with nearest expert actions -> training the final one-step policy. The only self-referential element is the standard reflow loop (Sec. III-A: 'Following [3], we train the reflow policy network using the optimal coupling (a_base_T, a_base_0) as the substitution of the former pairs'), but this is distillation, not circular: the final policy is evaluated in closed-loop task success, not on reconstructing those couplings, and the replacement actions come from external expert demonstrations. No fitted parameter is renamed as a prediction; the threshold delta is heuristic and unreported, but there is no evidence it encodes the final success rates. Two non-circular weaknesses are flagged: Sec. III-B asserts the selectively aligned coupling 'should maintain the property of optimal transport efficiency' without proof, and Algorithm 2's search is ambiguous between action-space and observation-space. Both are correctness/reproducibility concerns, not cases where the derivation reduces to its inputs by construction.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 0 invented entities

The central method has two hand-set quantities (delta and the coupling count N), and it depends on the density of expert demonstrations relative to the generated action distribution. No new physical or methodological entities are postulated.

free parameters (2)
  • delta (selective alignment threshold) = not reported
    Controls whether a generated action is replaced by a nearby expert action; chosen heuristically in Section III-B and used in Algorithm 2.
  • N = 10 couplings per action space = 10
    Number of noise-action couplings sampled for reflow/SeFA training; chosen to save generation and training time, but it affects the quality of the reflow policy and the final result.
axioms (3)
  • standard math Rectified flow reflow straightens sampling paths so a one-step Euler solve is accurate
    Invokes Liu et al. [3] reflow results to justify one-step inference; this is a background result the paper relies on.
  • domain assumption Expert demonstrations are dense enough that a nearest-neighbor expert action within delta is a valid correction
    The selective-alignment procedure only works if nearby expert actions exist for most generated actions; if the dataset is sparse, delta-based replacement would rarely trigger or would be unreliable.
  • domain assumption Reflow-induced observation-action inconsistency is the main cause of performance degradation
    The paper asserts that generated actions deviate from ground-truth actions and that errors accumulate, but this is not directly measured or proven.

pith-pipeline@v1.3.0-alltime-deepseek · 13191 in / 9745 out tokens · 101165 ms · 2026-08-03T22:46:15.127513+00:00 · methodology

0 comments
read the original abstract

Developing efficient and accurate visuomotor policies poses a central challenge in robotic imitation learning. While recent rectified flow approaches have advanced visuomotor policy learning, they suffer from a key limitation: After iterative distillation, generated actions may deviate from the ground-truth actions corresponding to the current visual observation, leading to accumulated error as the reflow process repeats and unstable task execution. We present Selective Flow Alignment (SeFA), an efficient and accurate visuomotor policy learning framework. SeFA resolves this challenge by a selective flow alignment strategy, which leverages expert demonstrations to selectively correct generated actions and restore consistency with observations, while preserving multimodality. This design introduces a consistency correction mechanism that ensures generated actions remain observation-aligned without sacrificing the efficiency of one-step flow inference. Extensive experiments across both simulated and real-world manipulation tasks show that SeFA Policy surpasses state-of-the-art diffusion-based and flow-based policies, achieving superior accuracy and robustness while reducing inference latency by over 98%. By unifying rectified flow efficiency with observation-consistent action generation, SeFA provides a scalable and dependable solution for real-time visuomotor policy learning. Code is available on https://github.com/RongXueZoe/SeFA.

Figures

Figures reproduced from arXiv: 2511.08583 by Jiageng Mao, Mingtong Zhang, Rong Xue, Yue Wang.

Figure 1
Figure 1. Figure 1: Selective Flow Alignment (SeFA) is a visual imitation learning algorithm that utilizes rectified flow with selective alignment, achieving superior effectiveness in diverse simulation and real-world tasks, with a significant inference acceleration. (a) Accuracy on various domains. (b) Sampling flow from noise to action of SeFA and Diffusion Policy. Abstract— Developing efficient and accurate visuomotor poli… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of SeFA. We train a visuomotor policy in an iterative manner to transport straight between noise distribution and target action space, hence enabling lightning one-step sampling during inference. The action flow is selectively aligned with observations, lowering the potential accumulated error brought by multiple reflows. models are computationally efficient because they can be sampled in a few or… view at source ↗
Figure 3
Figure 3. Figure 3: Sampling trajectories of SeFA-Policy at different stages. Randomly sampled pairs in (a) have crossing flows. Couplings in (b) have been rewired so they do not intersect with each other at the same denoising timestep. The trajectories in (c) and (d) are nearly straight. intentionally preserve this inherent multimodality to maintain the richness of the action distribution. After the selective alignment, we o… view at source ↗
Figure 4
Figure 4. Figure 4: Success Rates on Adroit (%). involve long-horizon sequential manipulation consisting of precise grasping and delicate wrist twisting, SeFA-Policy achieves 80% and 30% success rate, substantially outper￾forming Diffusion Policy’s 20% and 10%. This marked improvement underscores the efficacy of SeFA-Policy in gen￾erating consistent and accurate action sequences for multi￾stage, fine-grained continuous manipu… view at source ↗
Figure 7
Figure 7. Figure 7: Precise manipulation. SeFA successfully performs Rice Pouring, Knob Pull, Coffee Bean Sweeping, and Flower Insertion (from top to bottom). otherwise the drawer would stuck by friction. Also, Knob Pull includes precise grasping before the smooth pull action. In Coffee Bean Sweeping, the robot arm should apply a force onto the tabletop, which needs precise height control. VI. CONCLUSION In this work, we pres… view at source ↗
Figure 5
Figure 5. Figure 5: Grasping different objects with one policy. SeFA trained on the apple can generalize to other objects (cube, rubber duck) with similar sizes and locations [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Floating object manipulation. SeFA dynamically adjusts its action trajectory to approach and grab the moving rubber duck on the water, which demonstrates generalization ability to different object locations. drift, allowing successful execution even towards unfixed location. This highlights its suitability for real-world tasks involving dynamic objects. Precise manipulation. Flower Insertion and Rice Pouri… view at source ↗
Figure 8
Figure 8. Figure 8: Success Rate on 3D Adroit inputs (%). We evaluate 3 simulated tasks to show the accuracy variance of SeFA and 3D Diffusion Policy (3DP). TABLE VIII: Results on 3D inputs. We compare SeFA and 3D Diffusion Policy on 3D inputs from Adroit. Algorithm \ Task Adroit Average Hammer Door Pen SeFA 95±4 57±6 59±4 70±5 3D Diffusion Policy 83±12 53±12 50±15 62±13 REFERENCES [1] C. Chi, S. Feng, Y. Du, Z. Xu, E. Cousin… view at source ↗

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

45 extracted references · 7 linked inside Pith

  1. [1]

    Diffusion policy: Visuomotor policy learning via action diffusion,

    C. Chi, S. Feng, Y . Du, Z. Xu, E. Cousineau, B. Burchfiel, and S. Song, “Diffusion policy: Visuomotor policy learning via action diffusion,” arXiv preprint arXiv:2303.04137, 2023

    Pith/arXiv arXiv 2023

  2. [2]

    Adaflow: Imitation learn- ing with variance-adaptive flow-based policies,

    X. Hu, B. Liu, X. Liu, and Q. Liu, “Adaflow: Imitation learn- ing with variance-adaptive flow-based policies,”arXiv preprint arXiv:2402.04292, 2024

    Pith/arXiv arXiv 2024

  3. [3]

    Flow straight and fast: Learning to generate and transfer data with rectified flow,

    X. Liu, C. Gong, and qiang liu, “Flow straight and fast: Learning to generate and transfer data with rectified flow,” inThe Eleventh International Conference on Learning Representations, 2023. [Online]. Available: https://openreview.net/forum?id=XVjTT1nw5z

    2023

  4. [4]

    Planning with diffusion for flexible behavior synthesis,

    M. Janner, Y . Du, J. Tenenbaum, and S. Levine, “Planning with diffusion for flexible behavior synthesis,” inProceedings of the 39th International Conference on Machine Learning. PMLR, 2022, pp. 9902–9915, ISSN: 2640-3498. [Online]. Available: https://proceedings.mlr.press/v162/janner22a.html

    2022

  5. [5]

    Potential based diffusion motion planning,

    Y . Luo, C. Sun, J. B. Tenenbaum, and Y . Du, “Potential based diffusion motion planning,” inForty-first International Conference on Machine Learning, 2024

    2024

  6. [6]

    Motion plan- ning diffusion: Learning and planning of robot motions with diffusion models,

    J. Carvalho, A. Le, M. Baierl, D. Koert, and J. Peters, “Motion plan- ning diffusion: Learning and planning of robot motions with diffusion models,” inIEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2023

    2023

  7. [7]

    Edmp: Ensemble-of-costs-guided diffusion for motion planning,

    K. Saha, V . Mandadi, J. Reddy, A. Srikanth, A. Agarwal, B. Sen, A. Singh, and M. Krishna, “Edmp: Ensemble-of-costs-guided diffusion for motion planning,” in2024 IEEE International Conference on Robotics and Automation (ICRA), 2024, pp. 10 351–10 358

    2024

  8. [8]

    Diffusion-based generation, optimization, and planning in 3d scenes,

    S. Huang, Z. Wang, P. Li, B. Jia, T. Liu, Y . Zhu, W. Liang, and S.-C. Zhu, “Diffusion-based generation, optimization, and planning in 3d scenes,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023

    2023

  9. [9]

    Imitating human behaviour with diffusion models,

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

    2023

  10. [10]

    Diffusion policy: Visuomotor policy learning via action diffusion,

    C. Chi, S. Feng, Y . Du, Z. Xu, E. Cousineau, B. Burchfiel, and S. Song, “Diffusion policy: Visuomotor policy learning via action diffusion,” inRobotics: Science and Systems 2023. Robotics: Science and Systems Foundation. [Online]. Available: http://www.roboticsproceedings.org/rss19/p026.pdf

    2023

  11. [11]

    Scaling up and distilling down: Language-guided robot skill acquisition,

    H. Ha, P. Florence, and S. Song, “Scaling up and distilling down: Language-guided robot skill acquisition,” inProceedings of The 7th Conference on Robot Learning, ser. Proceedings of Machine Learning Research, J. Tan, M. Toussaint, and K. Darvish, Eds., vol

  12. [12]

    Chaineddiffuser: Unifying trajectory diffusion and keypose prediction for robotic manipulation,

    Z. Xian, N. Gkanatsios, T. Gervet, T.-W. Ke, and K. Fragkiadaki, “Chaineddiffuser: Unifying trajectory diffusion and keypose prediction for robotic manipulation,” inProceedings of The 7th Conference on Robot Learning, ser. Proceedings of Machine Learning Research, J. Tan, M. Toussaint, and K. Darvish, Eds., vol. 229. PMLR, 06–09 Nov 2023, pp. 2323–2339. [...

    2023

  13. [13]

    Crossway diffusion: Improving diffusion-based visuomotor policy via self-supervised learning

    X. Li, V . Belagali, J. Shang, and M. S. Ryoo, “Crossway diffusion: Improving diffusion-based visuomotor policy via self-supervised learning.” [Online]. Available: http://arxiv.org/abs/2307.01849

    Pith/arXiv arXiv

  14. [14]

    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

  15. [15]

    PoCo: Policy Composition from and for Heterogeneous Robot Learning,

    L. Wang, J. Zhao, Y . Du, E. Adelson, and R. Tedrake, “PoCo: Policy Composition from and for Heterogeneous Robot Learning,” inProceedings of Robotics: Science and Systems, Delft, Netherlands, July 2024

    2024

  16. [16]

    Don’t Start From Scratch: Behavioral Refinement via Interpolant-based Policy Diffusion,

    K. Chen, E. Lim, L. Kelvin, Y . Chen, and H. Soh, “Don’t Start From Scratch: Behavioral Refinement via Interpolant-based Policy Diffusion,” inProceedings of Robotics: Science and Systems, Delft, Netherlands, July 2024

    2024

  17. [17]

    Memory- consistent neural networks for imitation learning,

    K. Sridhar, S. Dutta, D. Jayaraman, J. Weimer, and I. Lee, “Memory- consistent neural networks for imitation learning,” 2023

    2023

  18. [18]

    ALOHA unleashed: A simple recipe for robot dexterity,

    T. Z. Zhao, J. Tompson, D. Driess, P. Florence, S. K. S. Ghasemipour, C. Finn, and A. Wahid, “ALOHA unleashed: A simple recipe for robot dexterity,” in8th Annual Conference on Robot Learning, 2024. [Online]. Available: https://openreview.net/forum?id=gvdXE7ikHI

    2024

  19. [19]

    Universal Manipulation Interface: In-The-Wild Robot Teaching Without In-The-Wild Robots,

    C. Chi, Z. Xu, C. Pan, E. Cousineau, B. Burchfiel, S. Feng, R. Tedrake, and S. Song, “Universal Manipulation Interface: In-The-Wild Robot Teaching Without In-The-Wild Robots,” inProceedings of Robotics: Science and Systems, Delft, Netherlands, July 2024

    2024

  20. [20]

    Goal-conditioned imitation learning using score-based diffusion policies,

    M. Reuss, M. Li, X. Jia, and R. Lioutikov, “Goal-conditioned imitation learning using score-based diffusion policies,” inRobotics: Science and Systems 2023. Robotics: Science and Systems Foundation. [Online]. Available: http://www.roboticsproceedings.org/rss19/p028. pdf

    2023

  21. [21]

    Multimodal Diffusion Transformer: Learning Versatile Behavior from Multimodal Goals,

    M. Reuss, ¨Omer Erdinc ¸ Ya˘gmurlu, F. Wenzel, and R. Lioutikov, “Multimodal Diffusion Transformer: Learning Versatile Behavior from Multimodal Goals,” inProceedings of Robotics: Science and Systems, Delft, Netherlands, July 2024

    2024

  22. [22]

    Playfusion: Skill acquisition via diffusion from language-annotated play,

    L. Chen, S. Bahl, and D. Pathak, “Playfusion: Skill acquisition via diffusion from language-annotated play,” inCoRL, 2023

    2023

  23. [23]

    Lad: Language control diffusion: efficiently scaling through space, time, and tasks,

    E. Zhang, Y . Lu, W. Wang, and A. Zhang, “Lad: Language control diffusion: efficiently scaling through space, time, and tasks,”arXiv preprint arXiv:2210.15629, 2023

    Pith/arXiv arXiv 2023

  24. [24]

    Se(3)-diffusionfields: Learning smooth cost functions for joint grasp and motion optimiza- tion through diffusion,

    J. Urain, N. Funk, J. Peters, and G. Chalvatzaki, “Se(3)-diffusionfields: Learning smooth cost functions for joint grasp and motion optimiza- tion through diffusion,”IEEE International Conference on Robotics and Automation (ICRA), 2023

    2023

  25. [25]

    Flow matching for generative modeling,

    Y . Lipman, R. T. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,” inThe Eleventh International Conference on Learning Representations, 2022

    2022

  26. [26]

    pi 0: A vision- language-action flow model for general robot control,

    K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichteret al., “pi 0: A vision- language-action flow model for general robot control,”arXiv preprint arXiv:2410.24164, 2024

    Pith/arXiv arXiv 2024

  27. [27]

    Elucidating the design space of diffusion-based generative models,

    T. Karras, M. Aittala, T. Aila, and S. Laine, “Elucidating the design space of diffusion-based generative models,” vol. 35, pp. 26 565– 26 577

  28. [28]

    Consistency models,

    Y . Song, P. Dhariwal, M. Chen, and I. Sutskever, “Consistency models,” inProceedings of the 40th International Conference on Machine Learning. PMLR, pp. 32 211–32 252, ISSN: 2640-3498. [Online]. Available: https://proceedings.mlr.press/v202/song23a.html

  29. [29]

    Denoising diffusion implicit models

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models.” [Online]. Available: https://openreview.net/forum? id=St1giarCHLP

  30. [30]

    Consistency trajectory models: 9 Learning probability flow ODE trajectory of diffusion,

    D. Kim, C.-H. Lai, W.-H. Liao, N. Murata, Y . Takida, T. Uesaka, Y . He, Y . Mitsufuji, and S. Ermon, “Consistency trajectory models: 9 Learning probability flow ODE trajectory of diffusion,” inThe Twelfth International Conference on Learning Representations, 2024. [Online]. Available: https://openreview.net/forum?id=ymjI8feDTD

    2024

  31. [31]

    Movement primitive diffusion: Learning gentle robotic manipulation of deformable objects,

    P. M. Scheikl, N. Schreiber, C. Haas, N. Freymuth, G. Neumann, R. Lioutikov, and F. Mathis-Ullrich, “Movement primitive diffusion: Learning gentle robotic manipulation of deformable objects,”IEEE Robotics and Automation Letters, vol. 9, no. 6, pp. 5338–5345, 2024

    2024

  32. [32]

    Streaming diffusion policy: Fast policy synthesis with variable noise diffusion models,

    S. H. Høeg, Y . Du, and O. Egeland, “Streaming diffusion policy: Fast policy synthesis with variable noise diffusion models,” 2024. [Online]. Available: https://arxiv.org/abs/2406.04806

    Pith/arXiv arXiv 2024

  33. [33]

    Rolling diffusion models,

    D. Ruhe, J. Heek, T. Salimans, and E. Hoogeboom, “Rolling diffusion models,” inProceedings of the 41st International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, R. Salakhutdinov, Z. Kolter, K. Heller, A. Weller, N. Oliver, J. Scarlett, and F. Berkenkamp, Eds., vol. 235. PMLR, 21–27 Jul 2024, pp. 42 818–42 835. [Online]. ...

    2024

  34. [34]

    Tedi: Temporally- entangled diffusion for long-term motion synthesis,

    Z. Zhang, R. Liu, R. Hanocka, and K. Aberman, “Tedi: Temporally- entangled diffusion for long-term motion synthesis,” inACM SIGGRAPH 2024 Conference Papers, ser. SIGGRAPH ’24. New York, NY , USA: Association for Computing Machinery, 2024. [Online]. Available: https://doi.org/10.1145/3641519.3657515

    arXiv 2024

  35. [35]

    Score-based generative modeling through stochastic differential equations,

    Y . Song, J. Sohl-dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochastic differential equations,” 2021. [Online]. Available: https://openreview. net/forum?id=PxTIG12RRHS

    2021

  36. [36]

    Learning complex dexterous manipulation with deep reinforcement learning and demonstrations,

    A. Rajeswaran, V . Kumar, A. Gupta, G. Vezzani, J. Schulman, E. Todorov, and S. Levine, “Learning complex dexterous manipulation with deep reinforcement learning and demonstrations,”arXiv, 2017

    2017

  37. [37]

    What matters in learning from offline human demonstrations for robot manipula- tion,

    A. Mandlekar, D. Xu, J. Wong, S. Nasiriany, C. Wang, R. Kulkarni, L. Fei-Fei, S. Savarese, Y . Zhu, and R. Mart´ın-Mart´ın, “What matters in learning from offline human demonstrations for robot manipula- tion,” in5th Annual Conference on Robot Learning, 2021

    2021

  38. [38]

    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,” inCoRL, 2020

    2020

  39. [39]

    Relay policy learning: Solving long horizon tasks via imitation and rein- forcement learning,

    A. Gupta, V . Kumar, C. Lynch, S. Levine, and K. Hausman, “Relay policy learning: Solving long horizon tasks via imitation and rein- forcement learning,”Conference on Robot Learning (CoRL), 2019

    2019

  40. [40]

    Implicit behavioral cloning,

    P. Florence, C. Lynch, A. Zeng, O. A. Ramirez, A. Wahid, L. Downs, A. Wong, J. Lee, I. Mordatch, and J. Tompson, “Implicit behavioral cloning,” inCoRL, 2022

    2022

  41. [41]

    Diffusion Policy: Visuomotor Policy Learning via Action Diffusion,

    C. Chi, S. Feng, Y . Du, Z. Xu, E. Cousineau, B. C. Burchfiel, and S. Song, “Diffusion Policy: Visuomotor Policy Learning via Action Diffusion,” inProceedings of Robotics: Science and Systems, Daegu, Republic of Korea, July 2023

    2023

  42. [42]

    Consistency policy: Accelerated visuomotor policies via consistency distillation,

    A. Prasad, K. Lin, J. Wu, L. Zhou, and J. Bohg, “Consistency policy: Accelerated visuomotor policies via consistency distillation,” 2024. [Online]. Available: http://arxiv.org/abs/2405.07503

    Pith/arXiv arXiv 2024

  43. [43]

    Scipy rk45 function,

    “Scipy rk45 function,” https://docs.scipy.org/doc/scipy/reference/ generated/scipy.integrate.RK45.html, accessed: 2022-08-19

    2022

  44. [44]

    Denoising diffusion implicit mod- els,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit mod- els,” inInternational Conference on Learning Representations, 2021. [Online]. Available: https://openreview.net/forum?id=St1giarCHLP 10

    2021

  45. [229]

    3766–3777

    PMLR, 06–09 Nov 2023, pp. 3766–3777. [Online]. Available: https://proceedings.mlr.press/v229/ha23a.html

    2023