Pith. sign in

REVIEW 3 major objections 5 minor 44 references

One forward pass can match 10-step generative robot policies when spatial, frequency, and multimodal errors are corrected together.

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 · grok-4.5

2026-07-11 23:24 UTC pith:YEC7ISQA

load-bearing objection Solid systems paper: three complementary regularizers make 1-NFE flow policies competitive with 10-step ManiFlow on 33 sim tasks plus real robots; the EMA-teacher residual is a real dependency but not a circular collapse. the 3 major comments →

arxiv 2607.03865 v1 pith:YEC7ISQA submitted 2026-07-04 cs.RO cs.AI

High-Fidelity One-Step Generative Visuomotor Policy via Recursive Correction, Frequency Consistency, and Contrastive Flow Matching

Yuran Chen , Xinye Cai , Zhonglin Gong , Yang Huang This is my paper
classification cs.RO cs.AI
keywords visuomotor policyflow matchingconsistency modelone-step generationrobot learningaction chunkingfrequency consistencycontrastive flow matching
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.

Multi-step diffusion and flow-matching policies give robots rich, multimodal actions, but the repeated sampling steps make closed-loop control too slow. Compressing that process into one big update usually produces three failures: spatial drift off the true action manifold, loss of high-frequency contact details, and averaged actions when several valid modes are entangled. This paper claims those three failures can be fixed by recursive residual correction, dual-timestep spectral consistency, and a margin-based contrastive objective, so that a single network evaluation recovers the fidelity of a 10-step baseline. Across bimanual, dexterous, and real-robot tasks the resulting one-step policy matches or beats strong multi-step methods while cutting inference to one function evaluation, making low-latency visuomotor control practical.

Core claim

The approximation gap of one-step generative visuomotor policies is not an inevitable cost of speed; it decomposes into spatial truncation, frequency distortion, and mode averaging, and these three failures can be closed simultaneously by recursive consistent action flow, dual-timestep frequency consistency, and contrastive flow matching, yielding success rates competitive with or better than 10-step baselines at 1 NFE.

What carries the argument

Recursive Consistent Action Flow (RCAF): an EMA teacher produces a long-span prediction and an N-step recursive tail; their residual is amortized into a clamped directional correction that the student matches, turning a large Euler step into a high-order corrected one-step velocity.

Load-bearing premise

The claim rests on the teacher’s short recursive trajectory (with fixed anchor ratio and step count) being a faithful enough proxy for the true curved action manifold that the residual correction remains informative rather than noise.

What would settle it

Train the full model and an otherwise identical ablated version that replaces RCAF’s recursive residual with ordinary mean-flow or plain consistency targets; if the 1-NFE success-rate gap to the 10-step baseline reappears on RoboTwin or Adroit, the recursive-correction claim fails.

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

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 / 5 minor

Summary. The paper proposes a one-step (1-NFE) generative visuomotor policy that addresses three failure modes of single-step acceleration—spatial truncation, frequency distortion, and mode averaging—via three modules: Recursive Consistent Action Flow (RCAF; residual correction from an EMA teacher’s long-span vs. N-step recursive trajectory, Eq. 5 and Alg. 1), Dual-Timestep Frequency Consistency (DTFC; adaptive DCT spectral consistency with focal and timestep-aware weights, Eq. 10), and Contrastive Flow Matching (CFM; margin-based repulsion with in-batch cyclic negatives, Eq. 11). The joint objective (Eq. 12) is trained on a DiT backbone and evaluated on 33 simulation tasks (RoboTwin, RoboTwin 2.0, Adroit, DexArt; 2D and 3D) plus real SO101 and UR7E platforms. The central claim is that the 1-NFE policy matches or exceeds strong 10-step baselines (ManiFlow, Diffusion Policy and 3D variants) in success rate while using a single forward pass.

Significance. If the empirical claim holds, the work is practically significant for high-frequency closed-loop robot control: it shows that carefully designed spatial residual correction, spectral consistency, and multimodal flow separation can close much of the quality gap between multi-step generative policies and true one-step inference. Strengths include broad multi-benchmark evaluation (33 sim tasks + two real platforms), clear incremental ablations isolating RCAF/DTFC/CFM (Tables III–VII), hyperparameter sweeps, qualitative flow-space and spectral analyses (Figs. 1, 6), and a control-theoretic fixed-point interpretation of RCAF (Appendix D). The contribution is primarily empirical and engineering-oriented rather than a new theoretical guarantee, but the combination is well-motivated for the robotics setting and the reported 1-NFE gains over 10-step ManiFlow are substantial if reproducible.

major comments (3)
  1. §IV-A, Eq. (5), Alg. 1 and Appendix D: RCAF’s residual ξ is computed entirely from the EMA teacher (v_long and the N-step D_rec). The fixed-point argument (Eqs. 13–17) assumes the teacher already approximates the multi-step manifold well enough that the high-gain feedback (k=1/ρ=5) is corrective. The paper does not report teacher–student alignment metrics, residual quality under early training or domain shift, or an ablation that freezes/degrades the teacher. Without such evidence, the strongest claim—that 1-NFE fidelity matches 10-step ManiFlow—rests on an unexamined dependency. A short teacher-quality diagnostic or frozen-teacher ablation would substantially strengthen the central result.
  2. §V-E and Table VIII: Real-world evaluation uses only 20 trials per setting. Several reported gains (e.g., SO101 Pos. Gen. 30 demos: 1/20 → 5/20; UR7E Stack: 4/20 → 7/20) are small absolute counts and could be sensitive to trial-to-trial variance. For a claim of reliable physical deployment, either more trials, confidence intervals, or a clearer statement of statistical uncertainty is needed so that the real-robot evidence can be weighed appropriately against the stronger simulation results.
  3. §V-C / Table I: Several baselines are re-implemented or marked with asterisks (e.g., 3D Flow Matching*). The manuscript should state more explicitly which numbers are taken from original papers versus re-runs under the authors’ protocol, and whether training budgets, observation encoders, and action horizons are matched. Without that, the large 1-NFE vs. 10-step gaps (especially on RoboTwin and Adroit pen) are harder to interpret as pure algorithmic gains.
minor comments (5)
  1. §IV-B / Alg. 2: The DTFC algorithm samples x_b = x_a + (t_b−t_a)v_true, which is consistent with the linear OT path only if v_true is constant; a one-sentence clarification that this is intentional under the OT assumption would help.
  2. Fig. 5g: The residual clamp ξ is said to drop near zero; relating this trajectory more explicitly to the theoretical fixed point in Appendix D would tighten the narrative.
  3. Notation: v_θ is overloaded for both instantaneous (Δt=0) and span-conditioned (Δt>0) predictions; a short notational distinction would reduce ambiguity in §IV and Alg. 4.
  4. Table II: Two tasks show ManiFlow slightly ahead (Click Bell, Lift Pot); a brief discussion of when the one-step method underperforms would balance the cross-domain narrative.
  5. Typos / polish: “PREPRINT SUBMITTED TO ARXIV” headers, occasional spacing issues (e.g., “First,Recursive”), and “π 0” formatting should be cleaned for the camera-ready version.

Circularity Check

0 steps flagged

No significant circularity: empirical method paper whose 1-NFE claims are tested on external task success rates, not forced by definition or self-citation chains.

full rationale

This is a standard robotic learning methods paper. RCAF, DTFC, and CFM are proposed training objectives (MSE residual correction from an EMA teacher, adaptive DCT spectral consistency, and a margin-based contrastive flow term). None of the central claims is a first-principles derivation that reduces by construction to its inputs. The load-bearing empirical claim—competitive or better success rates versus 10-step ManiFlow/Diffusion Policy on RoboTwin, RoboTwin 2.0, Adroit, DexArt, and real SO101/UR7E platforms—is measured by external task success, not by a quantity that is algebraically identical to a fitted parameter. The EMA-teacher residual in RCAF (Eq. 5, Algorithm 1) and the fixed-point analysis in Appendix D are ordinary self-distillation / consistency-model design and post-hoc dynamics analysis; they do not redefine the evaluation metric or smuggle a uniqueness theorem from the same authors. Hyperparameters (ρ, N, m, λ_freq, etc.) are chosen by ablation, not forced by a self-referential uniqueness claim. Citations for recursive consistency, mean flow, frequency consistency, and contrastive flow matching point to external prior work. No self-definitional loop, fitted-input-as-prediction, load-bearing self-citation uniqueness, or renaming of a known result as a new derivation is present. Concerns about teacher quality early in training are correctness/assumption risks, not circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 2 invented entities

Empirical robotics paper whose central claim rests on standard flow-matching and consistency assumptions plus a moderate number of hand-chosen hyperparameters that control residual scale, spectral focus, and contrastive margin. No new physical entities; the three named modules are algorithmic constructs whose value is measured by ablation success rates.

free parameters (5)
  • recursive steps N = 2
    Number of teacher sub-steps used to estimate D_rec; default N=2 chosen by ablation (Table IV).
  • anchor ratio ρ = 0.2
    Fraction of Δt used for the OT anchor step; default 0.2; ρ→0 causes divergence.
  • residual clamp γ = 5.0
    Bounds the correction magnitude ξ to prevent explosion; set to 5.0.
  • contrastive margin m = 0.4
    Static separation threshold in CFM; default 0.4 chosen over dynamic alternatives (Table VII).
  • λ_CFM / λ_freq / α / σ / E_warmup = 0.1 / 0.2 / 2.0 / 0.3 / 200
    Loss weights, focal exponent, Gaussian bandwidth and curriculum length; all set by hand or short sweeps and listed in Table IX.
axioms (4)
  • domain assumption Conditional OT path xt=(1−t)x0+tx1 yields constant target velocity x1−x0 that is a valid training signal for robotic action chunks.
    Standard flow-matching assumption (Eq. 2–3) imported from Lipman et al. and used throughout RCAF/DTFC/CFM.
  • domain assumption An EMA teacher provides a sufficiently accurate multi-step reference trajectory for residual correction.
    Core premise of RCAF (Section IV-A, Algorithm 1); if the teacher is poorly trained the residual ξ is noise.
  • domain assumption 1D Type-II DCT on the action-horizon dimension separates high- and low-frequency manipulation content useful for control.
    Invoked for DTFC (Eq. 7–10); spectral prior that earlier timesteps should emphasize low frequencies.
  • ad hoc to paper In-batch cyclic roll of target velocities supplies informative negative samples for mode separation.
    CFM sampling strategy (Algorithm 3); convenient but not theoretically guaranteed to cover true multimodal conflicts.
invented entities (2)
  • Recursive Consistent Action Flow residual ξ no independent evidence
    purpose: Explicit directional correction that amortizes long-span truncation error into the student velocity target.
    Defined in Eq. 5; no independent physical existence outside the training objective; value shown only by ablation gains.
  • Dual-Timestep Frequency Consistency adaptive weight ω_comb no independent evidence
    purpose: Focal × Gaussian spectral mask that protects high-frequency action details across timesteps.
    Constructed in Eq. 8–10; purely algorithmic; efficacy measured by success-rate lift over FreqPolicy/wavelet baselines.

pith-pipeline@v1.1.0-grok45 · 24131 in / 2999 out tokens · 26478 ms · 2026-07-11T23:24:27.522483+00:00 · methodology

0 comments
read the original abstract

Generative models such as diffusion and flow matching have advanced robotic visuomotor policies by modeling multimodal action distributions, but their multi-step sampling or ODE solving introduces inference latency. Existing one-step acceleration methods often compress the whole generation process into a single large update, leading to spatial deviation, frequency distortion, and mode averaging. This paper proposes a high-fidelity one-step generative visuomotor policy framework that addresses these issues with three complementary mechanisms. Recursive Consistent Action Flow (RCAF) uses recursive correction to compensate for spatial truncation errors and align one-step predictions with refined flow trajectories. Dual-Timestep Frequency Consistency (DTFC) preserves high-frequency manipulation details through adaptive spectral consistency across flow timesteps. Contrastive Flow Matching (CFM) separates entangled action flows with a margin-based repulsive objective, reducing ambiguous actions in multimodal manipulation. Experiments on RoboTwin, RoboTwin 2.0, Adroit, DexArt, and real-world robot platforms show that the proposed method achieves competitive or superior performance compared with strong 10-step generative policy baselines while requiring only one forward pass (1 NFE), enabling low-latency visuomotor control.

Figures

Figures reproduced from arXiv: 2607.03865 by Xinye Cai, Yang Huang, Yuran Chen, Zhonglin Gong.

Figure 1
Figure 1. Figure 1: Qualitative analysis of the single-step generation models. (a) Task sequence: Visual progression of the bimanual manipulation task. (b) Joint error [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed framework. (a) The multimodal Diffusion Transformer (DiT) policy network architecture. (b) Recursive Consistent Action [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of simulation benchmarks and real-world experiments. Left: the simulation evaluation suite includes 33 diverse tasks across DexArt, Adroit, RoboTwin 1.0, and RoboTwin 2.0, covering dexterous manipulation and complex bimanual coordination. Right: the real-world evaluation consists of multiple experiments on two physical 6-DoF robotic platforms: the SO101 arm with a global camera and a wrist-mounted… view at source ↗
Figure 4
Figure 4. Figure 4: Performance Comparison with SOTA Methods. Average success rates across five image-based manipulation benchmarks. Our single-step framework (NFE=1) achieves strong average performance while requiring an order of magnitude fewer inference steps than multi-step generative baselines. benchmark RoboTwin [25]. Efficacy of Core Components. Table III shows the incre￾mental contribution of each module. The 10-step … view at source ↗
Figure 5
Figure 5. Figure 5: Training dynamics and convergence analysis logged over 120k steps. The subplots systematically evaluate (a) the total loss, (b-d) the convergence [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comprehensive Qualitative Analysis and CFM Empirical Verification. (a-e) Action Generation Stability: (a) RGB observation sequence; (b) Absolute action error heatmaps; (c) 2D phase space action manifold; (d) Time-domain trajectories of grippers; (e) Frequency-domain velocity spectrum demonstrating high-frequency noise suppression. (f) Empirical Verification of CFM: Trajectories of 100 identical Gaussian no… 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

44 extracted references · 24 linked inside Pith

  1. [1]

    End-to-end training of deep visuomotor policies,

    S. Levine, C. Finn, T. Darrell, and P. Abbeel, “End-to-end training of deep visuomotor policies,”Journal of Machine Learning Research, vol. 17, no. 39, pp. 1–40, 2016

    2016

  2. [2]

    Imitation learning: Progress, taxonomies and challenges,

    B. Zheng, S. Verma, J. Zhou, I. W. Tsang, and F. Chen, “Imitation learning: Progress, taxonomies and challenges,”IEEE Transactions on Neural Networks and Learning Systems, vol. 35, no. 5, pp. 6322–6337, 2022

    2022

  3. [3]

    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

  4. [4]

    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 preprint arXiv:1709.10087, 2017

    Pith/arXiv arXiv 2017

  5. [5]

    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,”arXiv preprint arXiv:2403.03954, 2024

    Pith/arXiv arXiv 2024

  6. [6]

    Rdt-1b: a diffusion foundation model for bimanual manipula- tion,

    S. Liu, L. Wu, B. Li, H. Tan, H. Chen, Z. Wang, K. Xu, H. Su, and J. Zhu, “Rdt-1b: a diffusion foundation model for bimanual manipula- tion,”arXiv preprint arXiv:2410.07864, 2024

    Pith/arXiv arXiv 2024

  7. [7]

    π 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. Ichter,et al., “π 0: A Vision- Language-Action flow model for general robot control,”arXiv preprint arXiv:2410.24164, 2024

    Pith/arXiv arXiv 2024

  8. [8]

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

    T. Chen, Z. Chen, B. Chen, Z. Cai, Y . Liu, Z. Li, Q. Liang, X. Lin, Y . Ge, Z. Gu,et al., “Robotwin 2.0: A scalable data generator and benchmark with strong domain randomization for robust bimanual robotic manipulation,”arXiv preprint arXiv:2506.18088, 2025

    Pith/arXiv arXiv 2025

  9. [9]

    Visuomotor navigation for em- bodied robots with spatial memory and semantic reasoning cognition,

    Q. Liu, G. Wang, Z. Liu, and H. Wang, “Visuomotor navigation for em- bodied robots with spatial memory and semantic reasoning cognition,” IEEE Transactions on Neural Networks and Learning Systems, vol. 36, no. 5, pp. 9512–9523, 2024

    2024

  10. [10]

    Learning of long- horizon sparse-reward robotic manipulator tasks with base controllers,

    G. Wang, M. Xin, W. Wu, Z. Liu, and H. Wang, “Learning of long- horizon sparse-reward robotic manipulator tasks with base controllers,” IEEE Transactions on Neural Networks and Learning Systems, vol. 35, no. 3, pp. 4072–4081, 2022

    2022

  11. [11]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Advances in neural information processing systems, vol. 33, pp. 6840– 6851, 2020

    2020

  12. [12]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” arXiv preprint arXiv:2010.02502, 2020

    Pith/arXiv arXiv 2010

  13. [13]

    Flow matching for generative modeling,

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

    Pith/arXiv arXiv 2022

  14. [14]

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

    X. Liu, C. Gong, and Q. Liu, “Flow straight and fast: Learning to generate and transfer data with rectified flow,”arXiv preprint arXiv:2209.03003, 2022

    Pith/arXiv arXiv 2022

  15. [15]

    Learning robotic manipulation policies from point clouds with conditional flow matching,

    E. Chisari, N. Heppert, M. Argus, T. Welschehold, T. Brox, and A. Val- ada, “Learning robotic manipulation policies from point clouds with conditional flow matching,”arXiv preprint arXiv:2409.07343, 2024

    Pith/arXiv arXiv 2024

  16. [16]

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

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

    Pith/arXiv arXiv 2024

  17. [17]

    One-step diffusion policy: Fast visuomotor policies via diffusion distillation,

    Z. Wang, Z. Li, A. Mandlekar, Z. Xu, J. Fan, Y . Narang, L. Fan, Y . Zhu, Y . Balaji, M. Zhou,et al., “One-step diffusion policy: Fast visuomotor policies via diffusion distillation,”arXiv preprint arXiv:2410.21257, 2024

    Pith/arXiv arXiv 2024

  18. [18]

    Mp1: Meanflow tames policy learning in 1-step for robotic manipulation,

    J. Sheng, Z. Wang, P. Li, and M. Liu, “Mp1: Meanflow tames policy learning in 1-step for robotic manipulation,”arXiv preprint arXiv:2507.10543, 2025

    arXiv 2025

  19. [19]

    Dm1: Meanflow with dispersive regularization for 1-step robotic manipulation,

    G. Zou, H. Wang, H. Wu, Y . Qian, Y . Wang, and W. Li, “Dm1: Meanflow with dispersive regularization for 1-step robotic manipulation,”arXiv preprint arXiv:2510.07865, 2025

    arXiv 2025

  20. [20]

    Mean flow policy with instantaneous velocity constraint for one-step action generation,

    G. Zhan, L. Tao, P. Wang, Y . Wang, Y . Li, Y . Chen, M. Tomizuka, and S. E. Li, “Mean flow policy with instantaneous velocity constraint for one-step action generation,”arXiv preprint arXiv:2602.13810, 2026

    arXiv 2026

  21. [21]

    Maniflow: A general robot manipulation policy via consistency flow training,

    G. Yan, J. Zhu, Y . Deng, S. Yang, R.-Z. Qiu, X. Cheng, M. Mem- mel, R. Krishna, A. Goyal, X. Wang,et al., “Maniflow: A general robot manipulation policy via consistency flow training,”arXiv preprint arXiv:2509.01819, 2025

    Pith/arXiv arXiv 2025

  22. [22]

    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,”arXiv preprint arXiv:2405.07503, 2024

    Pith/arXiv arXiv 2024

  23. [23]

    Imitation learning policy based on multi-step consistent integration shortcut model,

    Y . Fang, X. Wang, X. Zhang, W. Xue, M. Zhang, S. Chen, and J. Zhao, “Imitation learning policy based on multi-step consistent integration shortcut model,”arXiv preprint arXiv:2510.19356, 2025

    arXiv 2025

  24. [24]

    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 Con- ference on Artificial Intelligence, vol. 39, pp. 14754–14762, 2025

    2025

  25. [25]

    Robotwin: Dual-arm robot benchmark with generative digital twins,

    Y . Mu, T. Chen, Z. Chen, S. Peng, Z. Lan, Z. Gao, Z. Liang, Q. Yu, Y . Zou, M. Xu,et al., “Robotwin: Dual-arm robot benchmark with generative digital twins,” inProceedings of the computer vision and pattern recognition conference, pp. 27649–27660, 2025

    2025

  26. [26]

    Dexart: Benchmarking generaliz- able dexterous manipulation with articulated objects,

    C. Bao, H. Xu, Y . Qin, and X. Wang, “Dexart: Benchmarking generaliz- able dexterous manipulation with articulated objects,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 21190–21200, 2023

    2023

  27. [27]

    3d diffuser ac- tor: Policy diffusion with 3d scene representations,

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

    Pith/arXiv arXiv 2024

  28. [28]

    3d flowmatch actor: Unified 3d policy for single-and dual-arm manipulation,

    N. Gkanatsios, J. Xu, M. Bronars, A. Mousavian, T.-W. Ke, and K. Fragkiadaki, “3d flowmatch actor: Unified 3d policy for single-and dual-arm manipulation,”arXiv preprint arXiv:2508.11002, 2025

    Pith/arXiv arXiv 2025

  29. [29]

    Consistency models,

    Y . Song, P. Dhariwal, M. Chen, and I. Sutskever, “Consistency models,” 2023

    2023

  30. [30]

    Simplifying, stabilizing and scaling continuous-time consistency models,

    C. Lu and Y . Song, “Simplifying, stabilizing and scaling continuous-time consistency models,”arXiv preprint arXiv:2410.11081, 2024

    Pith/arXiv arXiv 2024

  31. [31]

    Any-step generation via n-th order recursive consistent velocity field estimation,

    P. Sun and T. Lin, “Any-step generation via n-th order recursive consistent velocity field estimation,” inThe Fourteenth International Conference on Learning Representations

  32. [32]

    Consistency flow matching: Defining straight flows with velocity consistency,

    L. Yang, Z. Zhang, Z. Zhang, X. Liu, M. Xu, W. Zhang, C. Meng, S. Er- mon, and B. Cui, “Consistency flow matching: Defining straight flows with velocity consistency,”arXiv preprint arXiv:2407.02398, 2024

    Pith/arXiv arXiv 2024

  33. [33]

    Mean flows for one-step generative modeling,

    Z. Geng, M. Deng, X. Bai, J. Z. Kolter, and K. He, “Mean flows for one-step generative modeling,”arXiv preprint arXiv:2505.13447, 2025

    Pith/arXiv arXiv 2025

  34. [34]

    Improved mean flows: On the challenges of fastforward generative models,

    Z. Geng, Y . Lu, Z. Wu, E. Shechtman, J. Z. Kolter, and K. He, “Improved mean flows: On the challenges of fastforward generative models,”arXiv preprint arXiv:2512.02012, 2025

    Pith/arXiv arXiv 2025

  35. [35]

    Sparse diffusion policy: A sparse, reusable, and flexible policy for robot learning,

    Y . Wang, Y . Zhang, M. Huo, R. Tian, X. Zhang, Y . Xie, C. Xu, P. Ji, W. Zhan, M. Ding,et al., “Sparse diffusion policy: A sparse, reusable, and flexible policy for robot learning,”arXiv preprint arXiv:2407.01531, 2024

    Pith/arXiv arXiv 2024

  36. [36]

    Streaming flow policy: Simplifying diffusion/flow-matching policies by treating action trajectories as flow trajectories,

    S. Jiang, X. Fang, N. Roy, T. Lozano-P ´erez, L. P. Kaelbling, and S. Ancha, “Streaming flow policy: Simplifying diffusion/flow-matching policies by treating action trajectories as flow trajectories,”arXiv preprint arXiv:2505.21851, 2025

    arXiv 2025

  37. [37]

    Step: Warm-started visuomotor policies with spatiotemporal consistency prediction,

    J. Li, Y . Cong, Y . Wang, H. Xia, S. Huang, Y . Zhang, N. Xu, and G. Dai, “Step: Warm-started visuomotor policies with spatiotemporal consistency prediction,”arXiv preprint arXiv:2602.08245, 2026

    Pith/arXiv arXiv 2026

  38. [38]

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

    A. Sochopoulos, N. Malkin, N. Tsagkas, J. Moura, M. Gienger, and S. Vijayakumar, “Fast flow-based visuomotor policies via conditional optimal transport couplings,”arXiv preprint arXiv:2505.01179, 2025

    Pith/arXiv arXiv 2025

  39. [39]

    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,”arXiv preprint arXiv:2506.08822, 2025

    arXiv 2025

  40. [40]

    Scalable diffusion models with transformers,

    W. Peebles and S. Xie, “Scalable diffusion models with transformers,” inProceedings of the IEEE/CVF international conference on computer vision, pp. 4195–4205, 2023

    2023

  41. [41]

    One step diffusion via shortcut models,

    K. Frans, D. Hafner, S. Levine, and P. Abbeel, “One step diffusion via shortcut models,”arXiv preprint arXiv:2410.12557, 2024

    Pith/arXiv arXiv 2024

  42. [42]

    Contrastive flow matching,

    G. Stoica, V . Ramanujan, X. Fan, A. Farhadi, R. Krishna, and J. Hoff- man, “Contrastive flow matching,” inProceedings of the IEEE/CVF International Conference on Computer Vision, pp. 1185–1194, 2025

    2025

  43. [43]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE conference on computer vision and pattern recognition, pp. 770–778, 2016

    2016

  44. [44]

    Splitmeanflow: Interval splitting consistency in few-step generative modeling,

    Y . Guo, W. Wang, Z. Yuan, R. Cao, K. Chen, Z. Chen, Y . Huo, Y . Zhang, Y . Wang, S. Liu,et al., “Splitmeanflow: Interval splitting consistency in few-step generative modeling,”arXiv preprint arXiv:2507.16884, 2025

    Pith/arXiv arXiv 2025