Pith. sign in

REVIEW 3 major objections 5 minor 37 references

Robot policies that shorten action chunks when the observation is most sensitive succeed more often without extra computation.

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 14:17 UTC pith:O4WBJ64V

load-bearing objection Clean optimization-derived sensitivity metric for adaptive chunk horizons in diffusion policies, with matched-compute gains on sim/real; optimality claim rests on an unchecked power-law disturbance model that experiments never validate. the 3 major comments →

arxiv 2607.04739 v1 pith:O4WBJ64V submitted 2026-07-06 cs.RO

Spatial Attention: Adapting Execution Horizons for Diffusion Policies via Observation Sensitivity

Che-Sang Park , Junsu Ha , Jianlong Fu , Frank C. Park This is my paper
classification cs.RO
keywords diffusion policyaction chunkingexecution horizonSpatial Attentionlearning from demonstrationrobotic manipulationscore-based generative models
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.

Generative robot policies sample multi-step action chunks from a single observation and then execute a fixed-length prefix before looking again. That fixed length is a permanent trade-off: short chunks keep the robot responsive but waste computation; long chunks are cheap but blind to disturbances. This paper shows that the right length is not constant. It defines Spatial Attention as the expected squared size of the gradient of the action log-likelihood with respect to the observation—how much the policy’s action distribution would shift if the observation changed a little. Under a fixed total number of chunk samples, the length that keeps cumulative likelihood drop smallest shrinks exactly when Spatial Attention is large. A lightweight forecaster predicts those future values from the current observation and the just-sampled chunk, then cuts the chunk at the first moment the integrated rate exceeds a budget threshold. On both clean and deliberately disturbed manipulation tasks in simulation and on a real arm, the resulting adaptive policy raises success rates while the average number of samples stays the same.

Core claim

Under a fixed budget of chunk samplings, the execution horizon that minimizes the cumulative likelihood drop caused by observation disturbances decreases as Spatial Attention increases. Forecasting Spatial Attention along each sampled chunk and terminating when the integrated rate reaches a threshold therefore improves success rates while preserving the average execution horizon.

What carries the argument

Spatial Attention—the expected squared norm of the gradient of the action log-likelihood with respect to the observation—together with the optimal resampling rate ρ(t) ∝ Att(o(t))^{1/(2γ+1)} that solves the constrained minimization of cumulative likelihood drop.

Load-bearing premise

The paper models the observation disturbance that accumulates during open-loop execution as zero-mean isotropic Gaussian noise whose variance grows exactly as a power of elapsed time.

What would settle it

On a high-precision task such as Tool Hang, replace the power-law disturbance model with a non-Gaussian or non-power-law process; if the Spatial-Attention adaptive horizon then fails to raise success rate above the matched fixed-horizon baseline, the central claim is refuted.

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 addresses the fixed-execution-horizon trade-off in generative (diffusion/consistency) policies that sample action chunks. It defines Spatial Attention as Att(o) = E_a[||∇_o log π(a|o)||²], the expected squared sensitivity of the action log-likelihood to the observation. Under a fixed sampling budget N, it formulates minimization of cumulative expected log-likelihood drop caused by open-loop disturbances, approximates the drop via second-order Taylor expansion plus the Fisher-information identity, and obtains the optimal resampling rate ρ(t) ∝ Att(o(t))^{1/(2γ+1)}. At inference a sequence-to-sequence forecaster predicts future Att values from the current observation and sampled chunk; the execution horizon T_a is the first index at which the integrated rate exceeds a threshold C_att. Score networks (via Bayes decomposition) and a VAE latent for vision make Att computable without an explicit policy density. Experiments on Robomimic (standard and object-perturbation variants) and a real Franka task with a moving cube show higher success rates than fixed-horizon DDPM/DDIM/CP baselines while matching average T_a.

Significance. If the result holds, the work supplies a modular, architecture-agnostic criterion for adaptive chunk execution that is derived from an explicit optimization problem rather than a pure heuristic. The experimental protocol is careful: C_att is chosen so that mean T_a lies in [T_avg−0.5, T_avg+0.5], results are averaged over 20 checkpoints (4 seeds × 5 best), and both state and vision settings plus real-robot disturbance are reported. The real-robot lift of success rate from 0.42 to 0.92 at essentially identical average horizon is a concrete practical contribution. The framework applies to any chunk-generating policy and is therefore of interest beyond the diffusion models used for evaluation. The main theoretical claim is conditional on a modeling assumption (power-law isotropic Gaussian disturbance growth); when that assumption is accepted or the method is viewed as a principled heuristic, the paper advances the responsiveness–compute trade-off for generative robot policies.

major comments (3)
  1. [§III-A, Eq. (4); Abstract; Eqs. 8, 14] The optimality claim (Abstract; §III-B–C; Eqs. 8, 14) is derived under the modeling assumption that accumulated observation disturbance after open-loop time τ is zero-mean isotropic Gaussian with variance exactly ε²τ^{2γ} (Eq. 4). The subsequent Taylor + Fisher reduction yields D(o,τ)≈(1/2)Att(o)ε²τ^{2γ} (Eqs. 9–11), which is the sole justification for the continuous objective and the power-law rate. The experimental disturbances (orthogonal object translation once the gripper is near; constant-velocity camera-slider motion) are deterministic, structured, and non-Gaussian. The manuscript never checks whether residual observation error under these regimes grows as a power of open-loop time, nor whether the second-order approximation remains accurate. Consequently the claimed optimality of the Spatial-Attention schedule does not automatically transfer to the reported experiments. Either va
  2. [Eqs. (14), (20); §V experimental protocol] The free parameter γ enters the optimal rate exponent 1/(2γ+1) (Eq. 14) and the discrete accumulation rule (Eq. 20). The manuscript does not report the numerical value of γ used in any experiment, nor any sensitivity of success rate or realized T_a distribution to γ. Without this information the mapping from forecasted Att to T_a is not fully reproducible, and it is unclear how much of the reported gains depend on a particular choice of the disturbance-growth exponent versus the qualitative idea of shortening horizons in high-Att phases. Report γ (and the resulting effective exponent) for all tables, and preferably include a brief sensitivity sweep or an argument that results are stable over a plausible range of γ.
  3. [§II-B; Tables I–III] Related work (§II-B) criticizes Mixture of Horizons (MoH) as heuristic and notes that it was not evaluated under external disturbances, yet the experimental section contains no comparison to MoH or to any other adaptive-horizon baseline. Because the central empirical claim is that Spatial Attention improves success rates at matched average horizon, a direct head-to-head (or at least a discussion of why MoH is inapplicable to the disturbance settings) is needed to substantiate superiority over the closest published alternative. If computational or architectural reasons preclude a full re-implementation, state them and provide a controlled ablation that isolates the contribution of the Att-based schedule versus simpler adaptive rules (e.g., fixed short horizon near contact).
minor comments (5)
  1. [Fig. 2] Figure 2 caption and surrounding text usefully illustrate high-Att phases, but the quantitative scale of Att and the precise C_att threshold used for that rollout are not stated; adding them would make the figure self-contained.
  2. [§IV-B, Eq. (18)] The VAE latent proxy for vision (Eq. 18) is presented as practical; a short remark on whether Att computed in latent space correlates with Att computed (where feasible) in a lower-dimensional state space would strengthen the proxy claim.
  3. [Throughout; §IV-C heading] Notation: T_H, T_a, T_avg, T_max appear with slight typographic inconsistency (subscripts vs. plain); unify. Also “F orecasted” and similar spacing artifacts appear in headings.
  4. [§IV-C] The forecasting model Att_φ is a transformer trained offline; no quantitative forecast error (e.g., MAE of predicted vs. oracle Att along held-out trajectories) is reported. Even a brief number would help readers judge how much of the gain comes from accurate foresight versus the Att signal itself.
  5. [Table II] Table II omits the T_avg / steps columns present in Table I; for consistency and readability they should be restored or explicitly cross-referenced.

Circularity Check

0 steps flagged

No significant circularity; the optimality relation and adaptive-horizon rule are derived from an explicit approximate objective under a stated disturbance model, and C_att is used only to equalize average compute for fair comparison.

full rationale

The derivation chain is self-contained. Spatial Attention is defined directly as Att(o) ≜ E_a[‖∇_o log π(a|o)‖²] (Eq. 15). Under the modeling assumption δo ∼ N(0, ε² τ^{2γ} I) (Eq. 4), a second-order Taylor expansion plus the Fisher-information identity produces the closed-form drop D(o,τ) ≈ (1/2) Att(o) ε² τ^{2γ} (Eqs. 9–11). Substituting into the cumulative-drop objective and taking the continuum limit yields the tractable problem (Eq. 8). A standard Lagrangian then gives the pointwise stationarity condition ρ(t) ∝ Att(o(t))^{1/(2γ+1)} (Eq. 14). At inference the same quantity is forecasted by a sequence model trained on offline labels and the execution horizon is cut when the discrete integral of Att^{1/(2γ+1)} reaches a threshold C_att (Eq. 20). C_att is chosen solely so that the realized mean horizon lies inside [T_avg−0.5, T_avg+0.5], equalizing the sampling budget with the fixed-horizon baselines; it is not fitted to success rates. The success-rate gains reported in Tables I–III are therefore empirical outcomes, not tautological consequences of the fit. The sole self-citation ([14], coauthored by Park) appears only as motivational background for the phrase “minimum-attention control” and is not used to import uniqueness, an ansatz, or any load-bearing lemma. No equation reduces a claimed prediction to its own inputs by construction.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 2 invented entities

The central optimality claim rests on a small set of modeling choices (Gaussian power-law disturbance, second-order expansion, Fisher identity) that are standard in information geometry yet not independently verified for the robotic setting, plus two free scalars (γ, C_att) that control the schedule. Spatial Attention itself is an invented but operationally defined entity; the VAE latent and the forecasting transformer are auxiliary constructs required for vision and online use.

free parameters (3)
  • γ (disturbance growth exponent)
    Controls how fast open-loop disturbance variance grows with time; appears in the optimal exponent 1/(2γ+1) and is never fitted or ablated in the reported experiments.
  • C_att (attention accumulation threshold)
    Hand-tuned per task so that realized average execution horizon matches the fixed-horizon baseline; directly determines when the chunk is cut.
  • ε (noise scale)
    Overall scale of the Gaussian disturbance model; absorbed into the constant of the continuous objective but still part of the generative story for D(o,τ).
axioms (4)
  • ad hoc to paper Accumulated observation disturbance after open-loop interval τ is zero-mean isotropic Gaussian with variance ε²τ^{2γ}.
    Eq. 4; chosen for analytic convenience; no empirical validation against real robot noise or contact events.
  • standard math Second-order Taylor expansion of log π(a|o+δo) plus Fisher-information identity yields D(o,τ)≈(1/2)Att(o)ε²τ^{2γ}.
    Eqs. 9–11; standard local approximation; higher-order terms and non-Gaussian residuals are neglected.
  • domain assumption Att(o(·)) and ρ(·) are continuous, allowing the discrete bound (Eq. 12) to pass to the continuous objective (Eq. 8).
    Stated after Eq. 12; required for the pointwise stationarity argument that produces ρ∝Att^{1/(2γ+1)}.
  • domain assumption Latent-space gradient norm of a VAE encoder is a faithful proxy for observation-space sensitivity of the policy.
    Eq. 18; used for all vision experiments; no ablation against pixel-space or alternative encoders.
invented entities (2)
  • Spatial Attention Att(o)=E_a[‖∇_o log π(a|o)‖²] no independent evidence
    purpose: Scalar that quantifies policy sensitivity and directly sets the optimal resampling rate.
    Defined in Eq. 15; operational via score networks, but the name and the optimality link are introduced here; independent_evidence is false because the quantity is not measured outside the paper’s own success-rate tables.
  • Sequence-to-sequence Spatial Attention forecaster Att_ϕ no independent evidence
    purpose: Predicts future Att values from current observation and sampled action chunk so that the horizon can be chosen without generating future observations.
    Eq. 19; trained offline on labeled demonstration trajectories; no external validation of forecast accuracy.

pith-pipeline@v1.1.0-grok45 · 17865 in / 3215 out tokens · 29080 ms · 2026-07-11T14:17:13.445885+00:00 · methodology

0 comments
read the original abstract

Sampling action chunks via generative models has become a widely adopted methodology for robotic learning from demonstration. However, existing methods often struggle to balance responsiveness and computational cost because they execute each action chunk for a fixed execution horizon. In this paper, we adaptively adjust the execution horizon of sampled action chunks, balancing responsiveness and computational efficiency. We introduce Spatial Attention -- defined as the expected squared norm of the gradient of the action log-likelihood with respect to the observation -- which indicates the sensitivity of the policy's action distribution to variations in the observation. We show that, under a fixed budget of chunk samplings, the execution horizon that minimizes the cumulative likelihood drop induced by disturbances decreases as Spatial Attention increases. By forecasting future Spatial Attention values alongside the action chunk, our framework dynamically assigns shorter execution horizons to phases with high Spatial Attention, and longer horizons to phases with low Spatial Attention. Experiments on standard and perturbed tasks, in both simulation and on a real robot, show that our method significantly improves success rates over fixed-horizon baselines while maintaining the average execution horizon.

Figures

Figures reproduced from arXiv: 2607.04739 by Che-Sang Park, Frank C. Park, Jianlong Fu, Junsu Ha.

Figure 1
Figure 1. Figure 1: (a) Spatial Attention (SA) quantifies the sensitivity of the policy’s probability density function to variations in the observation. While a distribution [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Spatial Attention values during Tool Hang task. The top plot illustrates the evolution of Spatial Attention throughout the task, where green crosses mark the moments of observation and action chunk sampling. The frames below depict the physical states of the robot corresponding to these connected crosses. Notably, Spatial Attention increases during important manipulation phases, such as grasping ((a), (c))… view at source ↗
Figure 3
Figure 3. Figure 3: Robomimic Benchmark. We evaluate our method against the baseline policies on single-robot manipulation tasks from the Robomimic [15] benchmark. We categorize the Lift and Can tasks as “Simple” due to their lower precision requirements, while the Square and Tool Hang tasks are classified as “Complex” as they require precise manipulation. for each task. For the adaptive execution horizon with Spatial Attenti… view at source ↗
Figure 5
Figure 5. Figure 5: Real-World Experiment. (a) Hardware setup. The robot must grasp the red cube (white arrow) while a camera slider moves the cube laterally (red arrow). (b) The slider is placed at 5 different orientations to vary the direction of the cube’s motion relative to the robot. In this environment, we evaluate the success rate of our method and the baseline algorithms to analyze their responsiveness to external dis… 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

37 extracted references · 16 linked inside Pith

  1. [1]

    Diffusion policy: Visuomotor policy learning via ac- tion diffusion,

    C. Chi, Z. Xu, S. Feng, E. Cousineau, Y . Du, B. Burchfiel, R. Tedrake, and S. Song, “Diffusion policy: Visuomotor policy learning via ac- tion diffusion,”The International Journal of Robotics Research, p. 02783649241273668, 2023

    2023

  2. [2]

    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

  3. [3]

    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. Zhouet al., “One-step diffusion policy: Fast visuomotor policies via diffusion distillation,”arXiv preprint arXiv:2410.21257, 2024

    Pith/arXiv arXiv 2024

  4. [4]

    Learning fine- grained bimanual manipulation with low-cost hardware,

    T. Z. Zhao, V . Kumar, S. Levine, and C. Finn, “Learning fine- grained bimanual manipulation with low-cost hardware,”arXiv preprint arXiv:2304.13705, 2023

    Pith/arXiv arXiv 2023

  5. [5]

    Responsive noise-relaying diffusion policy: Responsive and efficient visuomotor control,

    Z. Chen, X. Yuan, T. Mu, and H. Su, “Responsive noise-relaying diffusion policy: Responsive and efficient visuomotor control,”arXiv preprint arXiv:2502.12724, 2025

    Pith/arXiv arXiv 2025

  6. [6]

    Reactive diffusion policy: Slow-fast visual-tactile policy learning for contact-rich manipulation,

    H. Xue, J. Ren, W. Chen, G. Zhang, Y . Fang, G. Gu, H. Xu, and C. Lu, “Reactive diffusion policy: Slow-fast visual-tactile policy learning for contact-rich manipulation,” inProceedings of Robotics: Science and Systems (RSS), 2025

    2025

  7. [7]

    Subconscious robotic imitation learning,

    J. Xie, Z. Wang, J. Tan, H. Lin, and X. Ma, “Subconscious robotic imitation learning,”arXiv preprint arXiv:2412.20368, 2024

    Pith/arXiv arXiv 2024

  8. [8]

    Openvla: An open- source vision-language-action model,

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. Foster, G. Lam, P. Sanketiet al., “Openvla: An open- source vision-language-action model,”arXiv preprint arXiv:2406.09246, 2024

    Pith/arXiv arXiv 2024

  9. [9]

    Real-time execution of action chunking flow policies,

    K. Black, M. Y . Galliker, and S. Levine, “Real-time execution of action chunking flow policies,”arXiv preprint arXiv:2506.07339, 2025

    Pith/arXiv arXiv 2025

  10. [10]

    Adaptive online replanning with diffusion models,

    S. Zhou, Y . Du, S. Zhang, M. Xu, Y . Shen, W. Xiao, D.-Y . Yeung, and C. Gan, “Adaptive online replanning with diffusion models,”Advances in Neural Information Processing Systems, vol. 36, pp. 44 000–44 016, 2023

    2023

  11. [11]

    Adaptive planning with generative models under uncertainty,

    P. Jutras-Dub ´e, R. Zhang, and A. Bera, “Adaptive planning with generative models under uncertainty,” in2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2024, pp. 7664–7670

    2024

  12. [12]

    Dynamic obstacle avoidance through uncertainty-based adaptive planning with diffusion,

    V . Punyamoorty, P. Jutras-Dub ´e, R. Zhang, V . Aggarwal, D. Conover, and A. Bera, “Dynamic obstacle avoidance through uncertainty-based adaptive planning with diffusion,”arXiv preprint arXiv:2409.16950, 2024

    Pith/arXiv arXiv 2024

  13. [13]

    Minimum attention control,

    W. Brockett, “Minimum attention control,” inProceedings of the 36th IEEE Conference on Decision and Control, vol. 3. IEEE, 1997, pp. 2628–2632

    1997

  14. [14]

    A minimum attention control law for ball catching,

    C. Jang, J.-e. Lee, S. Lee, and F. C. Park, “A minimum attention control law for ball catching,”Bioinspiration & Biomimetics, vol. 10, no. 5, p. 055008, 2015

    2015

  15. [15]

    What matters in learning from offline human demonstrations for robot manipulation,

    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 manipulation,” inConference on Robot Learning (CoRL), 2021

    2021

  16. [16]

    Action chunking and exploratory data collection yield exponential im- provements in behavior cloning for continuous control,

    T. T. Zhang, D. Pfrommer, C. Pan, N. Matni, and M. Simchowitz, “Action chunking and exploratory data collection yield exponential im- provements in behavior cloning for continuous control,”arXiv preprint arXiv:2507.09061, 2025

    arXiv 2025

  17. [17]

    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. Ichteret al., “π 0: A vision-language-action flow model for general robot control,”arXiv preprint arXiv:2410.24164, 2024

    Pith/arXiv arXiv 2024

  18. [18]

    GR00T N1: An open foundation model for generalist humanoid robots,

    NVIDIA, J. Bjorck, N. C. Fernando Casta ˜neda, X. Da, R. Ding, L. J. Fan, Y . Fang, D. Fox, F. Hu, S. Huang, J. Jang, Z. Jiang, J. Kautz, K. Kundalia, L. Lao, Z. Li, Z. Lin, K. Lin, G. Liu, E. Llontop, L. Magne, A. Mandlekar, A. Narayan, S. Nasiriany, S. Reed, Y . L. Tan, G. Wang, Z. Wang, J. Wang, Q. Wang, J. Xiang, Y . Xie, Y . Xu, Z. Xu, S. Ye, Z. Yu, ...

    2025

  19. [19]

    Planning with diffu- sion for flexible behavior synthesis,

    M. Janner, Y . Du, J. B. Tenenbaum, and S. Levine, “Planning with diffu- sion for flexible behavior synthesis,”arXiv preprint arXiv:2205.09991, 2022

    Pith/arXiv arXiv 2022

  20. [20]

    Is conditional generative modeling all you need for decision-making?

    A. Ajay, Y . Du, A. Gupta, J. Tenenbaum, T. Jaakkola, and P. Agrawal, “Is conditional generative modeling all you need for decision-making?” arXiv preprint arXiv:2211.15657, 2022

    Pith/arXiv arXiv 2022

  21. [21]

    Mixture of horizons in action chunking,

    D. Jing, G. Wang, J. Liu, W. Tang, Z. Sun, Y . Yao, Z. Wei, Y . Liu, Z. Lu, and M. Ding, “Mixture of horizons in action chunking,”arXiv preprint arXiv:2511.19433, 2025

    Pith/arXiv arXiv 2025

  22. [22]

    Generative modeling by estimating gradients of the data distribution,

    Y . Song and S. Ermon, “Generative modeling by estimating gradients of the data distribution,”Advances in neural information processing systems, vol. 32, 2019

    2019

  23. [23]

    Improved techniques for training score-based generative models,

    ——, “Improved techniques for training score-based generative models,” Advances in neural information processing systems, vol. 33, pp. 12 438– 12 448, 2020

    2020

  24. [24]

    Score-based generative modeling through stochastic differ- ential equations,

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochastic differ- ential equations,”arXiv preprint arXiv:2011.13456, 2020

    Pith/arXiv arXiv 2011

  25. [25]

    A connection between score matching and denoising au- toencoders,

    P. Vincent, “A connection between score matching and denoising au- toencoders,”Neural computation, vol. 23, no. 7, pp. 1661–1674, 2011

    2011

  26. [26]

    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

  27. [27]

    Deep unsupervised learning using nonequilibrium thermodynamics,

    J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli, “Deep unsupervised learning using nonequilibrium thermodynamics,” in International conference on machine learning. pmlr, 2015, pp. 2256– 2265

    2015

  28. [28]

    Diffusion models beat gans on image synthesis,

    P. Dhariwal and A. Nichol, “Diffusion models beat gans on image synthesis,”Advances in neural information processing systems, vol. 34, pp. 8780–8794, 2021

    2021

  29. [29]

    High- resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High- resolution image synthesis with latent diffusion models,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10 684–10 695

    2022

  30. [30]

    Auto-encoding variational bayes,

    D. P. Kingma and M. Welling, “Auto-encoding variational bayes,”arXiv preprint arXiv:1312.6114, 2013

    Pith/arXiv arXiv 2013

  31. [31]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,”Advances in neural information processing systems, vol. 30, 2017

    2017

  32. [32]

    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

  33. [33]

    Consistency trajectory models: 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: Learning probability flow ode trajectory of diffusion,”arXiv preprint arXiv:2310.02279, 2023

    Pith/arXiv arXiv 2023

  34. [34]

    Parallel sam- pling of diffusion models,

    A. Shih, S. Belkhale, S. Ermon, D. Sadigh, and N. Anari, “Parallel sam- pling of diffusion models,”Advances in Neural Information Processing Systems, vol. 36, pp. 4263–4276, 2023

    2023

  35. [35]

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

    B. Wen, W. Yang, J. Kautz, and S. Birchfield, “Foundationpose: Unified 6d pose estimation and tracking of novel objects,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 17 868–17 879

    2024

  36. [36]

    Fast policy synthesis with variable noise diffusion models,

    S. H. Høeg, Y . Du, and O. Egeland, “Fast policy synthesis with variable noise diffusion models,” in2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2025, pp. 4821–4828

    2025

  37. [37]

    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,” inConference on Robot Learning. PMLR, 2025, pp. 238–257

    2025