Pith. sign in

REVIEW 3 major objections 6 minor 54 references

SegDiff predicts continuous trajectories to the next keypose and uses DDIM inversion to keep action modes consistent in closed loop.

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

2026-07-14 07:33 UTC pith:TEYG5LFD

load-bearing objection Solid closed-loop recipe that anchors diffusion trajectories to keyposes and uses DDIM inversion for buffer checks; gains are real, premise is the usual PerAct keypose rule. the 3 major comments →

arxiv 2607.11027 v1 pith:TEYG5LFD submitted 2026-07-13 cs.RO

SegDiff: Segmented Trajectory Diffusion for Consistent and Adaptive Robot Manipulation

Haidong Cao , Wenjun Cao , Quanhao Li , Sicheng Xie , Zhiying Du , Jiaqi Leng , Zuxuan Wu , Yu-Gang Jiang This is my paper
classification cs.RO
keywords imitation learningdiffusion policykeyposesegmented trajectorytemporal ensemblingDDIM inversionrobot manipulationclosed-loop control
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.

Robot imitation learning has been split between short continuous action chunks that accumulate error and discrete keypose methods that need an external planner. SegDiff joins the two: it cuts demonstrations at keyposes (gripper open/close or near-zero end-effector speed) and trains a diffusion model to predict the full continuous path from the current state to the next keypose. At run time it executes only the first steps under receding-horizon control, keeps the rest in a buffer, and uses DDIM inversion of that buffer to generate a mode-consistent candidate. Two simple RMSE tests decide whether to blend the buffer, force the inverted mode, or discard the buffer entirely when the scene has changed. The result is long-horizon planning without open-loop planners, plus real-time reaction to multi-modal choices and dynamic disturbances. On RLBench, RoboMimic and five real robot tasks the method posts higher success rates than pure continuous diffusion and keypose baselines while adding only a few tens of milliseconds of latency.

Core claim

Segmented Trajectory Modeling (predicting the continuous trajectory that starts at the current state and ends at the next keypose) plus Dynamic Temporal Ensembling via DDIM inversion yields a closed-loop policy that simultaneously reduces compounding error, preserves action-mode consistency, and adapts to environmental change, producing higher success rates than continuous-prediction and keypose baselines on simulation and real-robot benchmarks.

What carries the argument

Segmented Trajectory Modeling together with Dynamic Temporal Ensembling: demonstrations are split at keyposes and the diffusion model is trained to map current observation and state onto an interpolated fixed-length trajectory that terminates at the next keypose; at inference DDIM inversion of the action buffer supplies a mode-consistent noise seed that is compared, via RMSE thresholds, against a fresh random-noise sample so the buffer can be refined, forced, or discarded.

Load-bearing premise

The method assumes that automatically extracted keyposes (gripper state change or near-zero velocity) are reliable segment anchors that absorb pose error and keep consecutive buffer and prediction trajectories aligned; if those points are poorly defined the training objective and the buffer-consistency test both fail.

What would settle it

Run the same multi-modal Package and dynamic Banana real-world tasks with deliberately noisy or missing keypose labels; if success rates fall to the level of ordinary Diffusion Policy and mode switches reappear, the claim that keypose anchors plus DDIM inversion are what produce the gains is falsified.

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 / 6 minor

Summary. SegDiff is a closed-loop visuomotor imitation policy that unifies continuous action prediction and keypose anchoring. Demonstrations are segmented by PerAct-style keyposes (gripper state change or near-zero end-effector velocity); a diffusion model is trained to predict the full continuous trajectory from an arbitrary current state to the next keypose (Segmented Trajectory Modeling, Eq. 5), with variable-length segments interpolated to fixed length L. At inference, Receding Horizon Control is combined with Dynamic Temporal Ensembling (DTE): DDIM inversion of the action buffer produces a prior-informed sample that is compared via RMSE to both the buffer and a random-noise sample; a data-driven threshold δ then decides whether to blend, preserve mode, or discard the buffer (Eq. 6). The method is evaluated on RLBench-10 (Table 1: 86.7% avg), full RLBench-60 (Fig. 5 / Table 11), RoboMimic data-efficiency curves (Fig. 6), and five real-world tasks (Table 2), with ablations of STM, DTE, L, η, and δ, plus a Lipschitz-style compounding-error analysis in Appendix A.

Significance. The paper targets well-recognized failure modes of behavior cloning—compounding error under short horizons, multi-modal action distributions, and open-loop keypose planners—and offers a single end-to-end diffusion architecture that keeps closed-loop control while retaining keypose structure. The empirical package is substantial: multi-seed RLBench-10, full RLBench-60, RoboMimic low-data regimes, and five real tasks covering position generalization, 6-DoF, multi-modality, dynamic perturbation, and precision. DTE via DDIM inversion is a concrete, implementable mechanism for mode-consistent temporal ensembling under changing observations, with measured real-robot overhead of ~5% (Table 6). Appendix A supplies an elementary but correctly scoped error bound contrasting O(L_a ε T²) sliding-window growth with O(L_a ε S) segment-local bounds. The primary limitation (dependence on pre-defined keyposes) is stated explicitly and flagged as future work. If the results hold under broader keypose definitions and larger real-world N, SegDiff is a useful practical advance for diffusion-based manipulation policies.

major comments (3)
  1. Appendix A (Eqs. 23–26) and the DTE alignment argument both treat keyposes as bottleneck states that approximately reset execution error (e_{k_m-1}≈0) and that share a common endpoint between buffer and new prediction. Keyposes are extracted solely by the PerAct rule (gripper open/close or velocity near zero; §3.2). The paper itself lists pre-defined keyposes as the primary limitation. There is no sensitivity study when keyposes are noisy, missing, or poorly timed (e.g., continuous contact tasks, multi-step assembly without clear velocity zeros). Because both the STM training objective and the buffer-validity check for DTE rest on this premise, the manuscript should either (i) quantify performance under perturbed/alternative keypose labels or (ii) strengthen the discussion of when the PerAct rule fails and how that would break DTE mode detection. Without that, the claimed O(L_a ε S) boun
  2. Real-world evaluation (Table 2) uses 10 trials per task and reports point success rates only. Package (multi-modality) and Banana (dynamic discard of buffer) are the primary empirical supports for DTE over vanilla temporal ensembling; Screw Cap is the only precision success (0.6) but with no baseline above 0. The N=10 design and lack of confidence intervals or multi-seed real runs make it hard to judge whether the large gaps (e.g., Banana 0.9 vs 0.2–0.4) are stable. Given that Table 3 shows DTE yields no gain on RLBench (as expected in monomodal sim), the load-bearing evidence for DTE is almost entirely these five small-N real tasks. Expanding trials, reporting binomial CIs, and/or adding one more dynamic/multi-modal task would substantially strengthen the central claim that DTE is necessary for real-world consistency and adaptation.
  3. Baseline reporting on RLBench-10 (Table 1) mixes numbers taken from CoA/Octo papers with a self-reproduction DP* that differs substantially from the published DP (e.g., Pick Up Cup 0.0 vs 76.0; Reach Target 8.0 vs 72.0). The paper correctly flags DP*, but the average comparison (SegDiff 86.7 vs DP 41.6 vs DP* 54.4) still leans on the weaker published DP for the headline gap. For fairness, either all continuous baselines should be re-run under the same observation stack, training budget, and temporal-ensembling settings used for SegDiff, or the main text should lead with DP* and treat published DP as secondary. As written, the magnitude of improvement over continuous diffusion is partly confounded by implementation variance.
minor comments (6)
  1. Fig. 1 and Fig. 2 use dense visual notation (action buffer, DTE paths) that is hard to parse in grayscale; a short legend or numbered callouts would help.
  2. Eq. 6 cases are clear, but the text never states how δ is computed at deployment when the training-set max RMSE is task-specific (Table 9 shows large per-task variation). A one-sentence protocol would remove ambiguity.
  3. Open Box is the only RLBench-10 task where SegDiff underperforms Octo (62.7 vs 96.0); a brief failure-mode note would balance the narrative.
  4. Notation: η is called both 'update factor' and (in one figure caption) 'fJ'; L is both interpolation length and Lipschitz constant in Appendix A—disambiguate.
  5. Related work could briefly situate SegDiff against concurrent consistency/streaming diffusion policies (e.g., Consistency Policy, Streaming Diffusion Policy) beyond the citations already present.
  6. Table 6 reports total latency ~656 ms; clarify whether this supports the claimed 'high-frequency' deployment or what control rate is actually used on the xArm7.

Circularity Check

0 steps flagged

No significant circularity: empirical method paper whose success rates are measured on held-out data and whose training objective is ordinary diffusion denoising on keypose-segmented trajectories.

full rationale

SegDiff is a closed-loop visuomotor policy that segments demonstrations by PerAct-style keyposes, trains a diffusion model (Eq. 5) to denoise fixed-length interpolated trajectories from arbitrary t to the next keypose, and at inference uses DDIM inversion of the action buffer plus RMSE thresholds (δ set to training-set max reconstruction error) for Dynamic Temporal Ensembling (Eq. 6). The central claims are empirical success-rate gains on RLBench-10/60, RoboMimic, and five real-world tasks versus ACT, DP, CoA, Octo, and AWE. These rates are obtained by executing the learned policy on held-out episodes and hardware; they are not algebraically forced by the training loss, the choice of L/η/δ, or any self-cited uniqueness theorem. Appendix A supplies a standard Lipschitz error-accumulation argument under the explicit modeling assumption that keyposes approximately reset pose error (e_{k_m-1}≈0); the resulting O(L_a ε S) bound is a consequence of that assumption, not a tautological re-labeling of the empirical results. Keypose extraction follows an external citation (PerAct), DDIM inversion is the standard Song et al. procedure, and ablations isolate STM versus DTE without circular dependence. No equation reduces a claimed prediction to a fitted input by construction, no load-bearing uniqueness result is imported from overlapping authors, and no known empirical pattern is merely renamed. The paper is therefore free of the six circularity patterns; score 0 is the correct outcome.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 1 invented entities

The central performance claim rests on standard imitation-learning and diffusion machinery plus a small set of design choices (keypose definition, fixed interpolation length, DTE thresholds and blend factor). No new physical entities are postulated; free parameters are ordinary hyper-parameters selected on validation or by training-set statistics.

free parameters (5)
  • Interpolation length L
    All variable-length segments are cubic/SLERP-interpolated to fixed length L; peak performance reported at L=25 (Fig. 9). Directly shapes the training target and buffer size.
  • Update factor η
    Blend weight between new prediction and action buffer in Eq. 6; default 0.25. Ablation shows η=0 and η=1 both degrade success.
  • Mode/validity threshold δ
    RMSE threshold deciding buffer validity and mode consistency; set to max (or p99/p95) DDIM reconstruction error on the training set. Controls when the buffer is discarded.
  • Number of executed actions n_exec and observation history h
    Receding-horizon execution length and history length; fixed to 1 in the reported tables but required for the RHC + buffer design.
  • Diffusion / inversion step counts (SS, IS)
    Sampling steps and DDIM inversion steps (default SS=50, IS=10); trade accuracy for latency (Table 5).
axioms (4)
  • domain assumption Keyposes are the timesteps where gripper state changes or end-effector velocity approaches zero (PerAct rule), and the final action is always a keypose.
    Section 3.2; used both to build training segments and to claim that keyposes absorb pose error.
  • standard math System dynamics and expert policy are Lipschitz continuous, and the learned policy has uniformly bounded error ε (Appendix A).
    Required for the compounding-error bound that justifies STM; never verified numerically on the real robot.
  • ad hoc to paper DDIM inversion of a buffered trajectory yields a noise vector that, when re-denoised under a new observation, stays in the same action mode if and only if the buffer remains feasible.
    Core of Dynamic Temporal Ensembling (Sec. 3.3, Fig. 3); supported by qualitative diagrams and ablations but not by a formal mode-separation theorem.
  • domain assumption Offline imitation from fixed expert datasets is sufficient; no online interaction is required.
    Stated in Sec. 3.1; standard OIL premise.
invented entities (1)
  • Dynamic Temporal Ensembling (DTE) decision rule no independent evidence
    purpose: Uses two RMSE metrics (d_valid, d_mode) and threshold δ to choose among random sample, inverted-noise sample, or buffer blend (Eq. 6).
    New control logic introduced by the paper; independent evidence is only the empirical ablations, not an external physical measurement.

pith-pipeline@v1.1.0-grok45 · 26136 in / 3485 out tokens · 31672 ms · 2026-07-14T07:33:02.082275+00:00 · methodology

0 comments
read the original abstract

Imitation learning enables robots to acquire manipulation skills from demonstrations by mapping observations to actions. Existing approaches predict either short-horizon continuous action sequences or discrete keyposes. However, continuous prediction methods suffer from compounding errors due to short prediction horizons and struggle with multi-modal action distributions, whereas keypose-based methods necessitate an external planner, constraining real-time applicability. To address these challenges, we introduce SegDiff, a closed-loop visuomotor policy that integrates the strengths of both paradigms. SegDiff decomposes demonstrations into motion segments between keyposes and learns to predict the continuous trajectory from the current state to the next keypose, enabling long-horizon prediction with real-time refinement. Furthermore, we leverage the capability of diffusion models and DDIM inversion to propose a Dynamic Temporal Ensembling mechanism, which allows the policy to efficiently respond to dynamic environments and mitigate discontinuities caused by inconsistent multi-modal sampling. SegDiff demonstrates significant performance gains over existing approaches across various simulated and real-world scenarios, indicating its strong ability to reason over extended temporal dependencies while maintaining real-time adaptability and control stability.

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

54 extracted references · 11 linked inside Pith

  1. [1]

    Sail: Faster-than-demonstration execution of imitation learning policies

    Nadun Ranawaka Arachchige, Zhenyang Chen, Wonsuhk Jung, Woo Chul Shin, Rohan Bansal, Pierre Barroso, Yu Hang He, Yingyan Celine Lin, Benjamin Joffe, Shreyas Kousik, et al. Sail: Faster-than-demonstration execution of imitation learning policies. InConferenceonRobotLearning, pages 721–749. PMLR, 2025

    2025

  2. [2]

    A framework for behavioural cloning

    Michael Bain and Claude Sammut. A framework for behavioural cloning. InMachineintelligence15, pages 103–129, 1995

    1995

  3. [3]

    Hydra: Hybrid robot actions for imitation learning

    Suneel Belkhale, Yuchen Cui, and Dorsa Sadigh. Hydra: Hybrid robot actions for imitation learning. InConference on RobotLearning, pages 2113–2133. PMLR, 2023

    2023

  4. [4]

    Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al.\𝜋0: A vision-language-action flow model for general robot control.arXiv preprint arXiv:2410.24164, 2024

    Pith/arXiv arXiv 2024

  5. [5]

    Real-time execution of action chunking flow policies.arXiv preprint arXiv:2506.07339, 2025

    Kevin Black, Manuel Y Galliker, and Sergey Levine. Real-time execution of action chunking flow policies.arXiv preprint arXiv:2506.07339, 2025

    Pith/arXiv arXiv 2025

  6. [6]

    Polarnet: 3d point clouds for language-guided robotic manipulation

    Shizhe Chen, Ricardo Garcia, Cordelia Schmid, and Ivan Laptev. Polarnet: 3d point clouds for language-guided robotic manipulation. In7th Conferenceon RobotLearning (CoRL2023), 2023

    2023

  7. [7]

    Bail: Best-action imitation learning for batch deep reinforcement learning.AdvancesinNeuralInformationProcessingSystems, 33:18353–18363, 2020

    Xinyue Chen, Zijian Zhou, Zheng Wang, Che Wang, Yanqiu Wu, and Keith Ross. Bail: Best-action imitation learning for batch deep reinforcement learning.AdvancesinNeuralInformationProcessingSystems, 33:18353–18363, 2020

    2020

  8. [8]

    Responsive noise-relaying diffusion policy: Responsive and efficient visuomotor control.arXivpreprintarXiv:2502.12724, 2025

    Zhuoqun Chen, Xiu Yuan, Tongzhou Mu, and Hao Su. Responsive noise-relaying diffusion policy: Responsive and efficient visuomotor control.arXivpreprintarXiv:2502.12724, 2025

    Pith/arXiv arXiv 2025

  9. [9]

    Diffusion policy: Visuomotor policy learning via action diffusion.The International Journal of RoboticsResearch, 44(10-11):1684–1704, 2025

    Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion.The International Journal of RoboticsResearch, 44(10-11):1684–1704, 2025

    2025

  10. [10]

    Real-time iteration scheme for diffusion policy.arXiv preprint arXiv:2508.05396, 2025

    Yufei Duan, Hang Yin, and Danica Kragic. Real-time iteration scheme for diffusion policy.arXiv preprint arXiv:2508.05396, 2025

    Pith/arXiv arXiv 2025

  11. [11]

    Sam2act: Integrating visual foundation model with a memory architecture for robotic manipulation

    Haoquan Fang, Markus Grotz, Wilbert Pumacay, Yi Ru Wang, Dieter Fox, Ranjay Krishna, and Jiafei Duan. Sam2act: Integrating visual foundation model with a memory architecture for robotic manipulation. InInternational Conferenceon MachineLearning, pages 15925–15942. PMLR, 2025

    2025

  12. [12]

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

    Theophile Gervet, Zhou Xian, Nikolaos Gkanatsios, and Katerina Fragkiadaki. Act3d: 3d feature field transformers for multi-task robotic manipulation. InConferenceon RobotLearning, pages 3949–3965. PMLR, 2023

    2023

  13. [13]

    Rvt: Robotic view transformer for 3d object manipulation

    Ankit Goyal, Jie Xu, Yijie Guo, Valts Blukis, Yu-Wei Chao, and Dieter Fox. Rvt: Robotic view transformer for 3d object manipulation. InConferenceonRobotLearning, pages 694–710. PMLR, 2023

    2023

  14. [14]

    Rvt2: Learning precise manipulation from few demonstrations.RSS, 2024

    Ankit Goyal, Valts Blukis, Jie Xu, Yijie Guo, Yu-Wei Chao, and Dieter Fox. Rvt2: Learning precise manipulation from few demonstrations.RSS, 2024

    2024

  15. [15]

    Denoisingdiffusionprobabilisticmodels

    JonathanHo,AjayJain,andPieterAbbeel. Denoisingdiffusionprobabilisticmodels. Advancesinneuralinformation processing systems, 33:6840–6851, 2020

    2020

  16. [16]

    Streaming diffusion policy: Fast policy synthesis with variable noise diffusion models.arXivpreprint arXiv:2406.04806, 2024

    Sigmund H Høeg, Yilun Du, and Olav Egeland. Streaming diffusion policy: Fast policy synthesis with variable noise diffusion models.arXivpreprint arXiv:2406.04806, 2024

    Pith/arXiv arXiv 2024

  17. [17]

    Video prediction policy: A generalist robot policy with predictive visual representations

    YuchengHu,YanjiangGuo,PengchaoWang,XiaoyuChen,Yen-JenWang,JiankeZhang,KoushilSreenath,Chaochao Lu, and Jianyu Chen. Video prediction policy: A generalist robot policy with predictive visual representations. In International Conferenceon MachineLearning, pages 24328–24346. PMLR, 2025

    2025

  18. [18]

    Q-attention: Enablingefficientlearningforvision-basedroboticmanipulation

    StephenJamesandAndrewJDavison. Q-attention: Enablingefficientlearningforvision-basedroboticmanipulation. IEEE Roboticsand AutomationLetters, 7(2):1612–1619, 2022

    2022

  19. [19]

    Rlbench: The robot learning benchmark & learning environment.IEEE Roboticsand AutomationLetters, 5(2):3019–3026, 2020

    Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J Davison. Rlbench: The robot learning benchmark & learning environment.IEEE Roboticsand AutomationLetters, 5(2):3019–3026, 2020. 13

    2020

  20. [20]

    Coarse-to-fine q-attention: Efficient learning forvisualroboticmanipulationviadiscretisation

    Stephen James, Kentaro Wada, Tristan Laidlow, and Andrew J Davison. Coarse-to-fine q-attention: Efficient learning forvisualroboticmanipulationviadiscretisation. In ProceedingsoftheIEEE/CVFConferenceonComputerVision and PatternRecognition, pages 13739–13748, 2022

    2022

  21. [21]

    Sampling-based algorithms for optimal motion planning.Theinternational journal ofroboticsresearch, 30(7):846–894, 2011

    Sertac Karaman and Emilio Frazzoli. Sampling-based algorithms for optimal motion planning.Theinternational journal ofroboticsresearch, 30(7):846–894, 2011

    2011

  22. [22]

    Anytime motion planning using the rrt

    Sertac Karaman, Matthew R Walter, Alejandro Perez, Emilio Frazzoli, and Seth Teller. Anytime motion planning using the rrt. In2011 IEEEinternational conferenceon roboticsand automation, pages 1478–1483. ieee, 2011

    2011

  23. [23]

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

    Tsung-Wei Ke, Nikolaos Gkanatsios, and Katerina Fragkiadaki. 3d diffuser actor: Policy diffusion with 3d scene representations. InConferenceonRobotLearning, pages 1949–1974. PMLR, 2025

    1949

  24. [24]

    Fine-tuning vision-language-action models: Optimizing speed and success

    Moo Jin Kim, Chelsea Finn, and Percy Liang. Fine-tuning vision-language-action models: Optimizing speed and success. arXivpreprintarXiv:2502.19645, 2025

    Pith/arXiv arXiv 2025

  25. [25]

    Rrt-connect: An efficient approach to single-query path planning

    James J Kuffner and Steven M LaValle. Rrt-connect: An efficient approach to single-query path planning. In Proceedings 2000 ICRA. Millennium conference. IEEE international conference on robotics and automation. Symposiaproceedings (Cat.No.00CH37065), volume 2, pages 995–1001. IEEE, 2000

    2000

  26. [26]

    Rapidly-exploring random trees: A new tool for path planning.ResearchReport9811, 1998

    Steven LaValle. Rapidly-exploring random trees: A new tool for path planning.ResearchReport9811, 1998

    1998

  27. [27]

    Curriculum offline imitating learning.AdvancesinNeuralInformationProcessingSystems, 34:6266–6277, 2021

    Minghuan Liu, Hanye Zhao, Zhengyu Yang, Jian Shen, Weinan Zhang, Li Zhao, and Tie-Yan Liu. Curriculum offline imitating learning.AdvancesinNeuralInformationProcessingSystems, 34:6266–6277, 2021

    2021

  28. [28]

    Rdt-1b: a diffusion foundation model for bimanual manipulation

    Songming Liu, Lingxuan Wu, Bangguo Li, Hengkai Tan, Huayu Chen, Zhengyi Wang, Ke Xu, Hang Su, and Jun Zhu. Rdt-1b: a diffusion foundation model for bimanual manipulation. InInternational Conferenceon Learning Representations, volume 2025, pages 29982–30009, 2025

    2025

  29. [29]

    Manicm: Real-time 3d diffusion policy via consistency model for robotic manipulation.arXivpreprintarXiv:2406.01586, 2024

    Guanxing Lu, Zifeng Gao, Tianxing Chen, Wenxun Dai, Ziwei Wang, Wenbo Ding, and Yansong Tang. Manicm: Real-time 3d diffusion policy via consistency model for robotic manipulation.arXivpreprintarXiv:2406.01586, 2024

    Pith/arXiv arXiv 2024

  30. [30]

    What matters in learning from offline human demonstrations for robot manipulation

    AjayMandlekar,DanfeiXu,JosiahWong,SoroushNasiriany,ChenWang,RohunKulkarni,LiFei-Fei,SilvioSavarese, Yuke Zhu, and Roberto Martín-Martín. What matters in learning from offline human demonstrations for robot manipulation. InConferenceon RobotLearning, pages 1678–1690. PMLR, 2022

    2022

  31. [31]

    Receding horizon control of nonlinear systems

    David Q Mayne and Hannah Michalska. Receding horizon control of nonlinear systems. InProceedingsofthe 27th IEEE Conferenceon Decisionand Control, pages 464–465. IEEE, 1988

    1988

  32. [32]

    An algorithmic perspective on imitation learning.Foundationsand Trends®inRobotics, 7(1-2):1–179, 2018

    Takayuki Osa, Joni Pajarinen, Gerhard Neumann, J Andrew Bagnell, Pieter Abbeel, Jan Peters, et al. An algorithmic perspective on imitation learning.Foundationsand Trends®inRobotics, 7(1-2):1–179, 2018

    2018

  33. [33]

    Consistencypolicy: Acceleratedvisuomotor policies via consistency distillation.arXivpreprintarXiv:2405.07503, 2024

    AadityaPrasad, KevinLin, JimmyWu, LinqiZhou, andJeannetteBohg. Consistencypolicy: Acceleratedvisuomotor policies via consistency distillation.arXivpreprintarXiv:2405.07503, 2024

    Pith/arXiv arXiv 2024

  34. [34]

    V-rep: A versatile and scalable robot simulation framework

    Eric Rohmer, Surya PN Singh, and Marc Freese. V-rep: A versatile and scalable robot simulation framework. In2013 IEEE/RSJinternational conferenceonintelligentrobotsandsystems, pages 1321–1326. IEEE, 2013

    2013

  35. [35]

    Learning from demonstration.Advancesinneuralinformationprocessingsystems, 9, 1996

    Stefan Schaal. Learning from demonstration.Advancesinneuralinformationprocessingsystems, 9, 1996

    1996

  36. [36]

    Is imitation learning the route to humanoid robots?Trendsincognitivesciences, 3(6):233–242, 1999

    Stefan Schaal. Is imitation learning the route to humanoid robots?Trendsincognitivesciences, 3(6):233–242, 1999

    1999

  37. [37]

    Waypoint-based imitation learning for robotic manipulation

    Lucy Xiaoyang Shi, Archit Sharma, Tony Z Zhao, and Chelsea Finn. Waypoint-based imitation learning for robotic manipulation. InConferenceon RobotLearning, pages 2195–2209. PMLR, 2023

    2023

  38. [38]

    Perceiver-actor: A multi-task transformer for robotic manipulation

    Mohit Shridhar, Lucas Manuelli, and Dieter Fox. Perceiver-actor: A multi-task transformer for robotic manipulation. In Conferenceon RobotLearning, pages 785–799. PMLR, 2023

    2023

  39. [39]

    Generative image as action models

    Mohit Shridhar, Yat Long Lo, and Stephen James. Generative image as action models. InConference on Robot Learning, pages 2429–2455. PMLR, 2025

    2025

  40. [40]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. InInternational Conference on Learning Representations, 2021. 14

    2021

  41. [41]

    What’s the move? hybrid imitation learning via salient points

    Priya Sundaresan, Hengyuan Hu, Quan Vuong, Jeannette Bohg, and Dorsa Sadigh. What’s the move? hybrid imitation learning via salient points. InInternational Conferenceon Learning Representations, volume 2025, pages 51806–51821, 2025

    2025

  42. [42]

    Octo: An open-source generalist robot policy.arXiv preprint arXiv:2405.12213, 2024

    Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joey Hejna, Tobias Kreiman, Charles Xu, et al. Octo: An open-source generalist robot policy.arXiv preprint arXiv:2405.12213, 2024

    Pith/arXiv arXiv 2024

  43. [43]

    Hierarchical diffusion policy: manipulation trajectory generation via contact guidance.IEEETransactionson Robotics, 2025

    Dexin Wang, Chunsheng Liu, Faliang Chang, and Yichen Xu. Hierarchical diffusion policy: manipulation trajectory generation via contact guidance.IEEETransactionson Robotics, 2025

    2025

  44. [44]

    Group normalization

    Yuxin Wu and Kaiming He. Group normalization. InProceedings of the European conferenceon computervision (ECCV), pages 3–19, 2018

    2018

  45. [45]

    Chaineddiffuser: Unifying trajectory diffusion and keypose prediction for robotic manipulation

    Zhou Xian, Nikolaos Gkanatsios, Theophile Gervet, Tsung-Wei Ke, and Katerina Fragkiadaki. Chaineddiffuser: Unifying trajectory diffusion and keypose prediction for robotic manipulation. In7th AnnualConferenceonRobot Learning, 2023

    2023

  46. [46]

    Human2robot: Learning robot actions from paired human-robot videos

    Sicheng Xie, Haidong Cao, Zejia Weng, Zhen Xing, Haoran Chen, Shiwei Shen, Jiaqi Leng, Zuxuan Wu, and Yu-Gang Jiang. Human2robot: Learning robot actions from paired human-robot videos. InProceedings of the AAAI Conferenceon ArtificialIntelligence, volume 40, pages 11078–11086, 2026

    2026

  47. [47]

    Bikc: Keypose-conditioned consistency policy for bimanual robotic manipulation.arXivpreprintarXiv:2406.10093, 2024

    Dongjie Yu, Hang Xu, Yizhou Chen, Yi Ren, and Jia Pan. Bikc: Keypose-conditioned consistency policy for bimanual robotic manipulation.arXivpreprintarXiv:2406.10093, 2024

    Pith/arXiv arXiv 2024

  48. [48]

    Asurveyofimitationlearning: Algorithms, recent developments, and challenges.IEEE Transactionson Cybernetics, 2024

    MaryamZare,ParhamMKebria,AbbasKhosravi,andSaeidNahavandi. Asurveyofimitationlearning: Algorithms, recent developments, and challenges.IEEE Transactionson Cybernetics, 2024

    2024

  49. [49]

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

    Yanjie Ze, Ge Yan, Yueh-Hua Wu, Annabella Macaluso, Yuying Ge, Jianglong Ye, Nicklas Hansen, Li Erran Li, and Xiaolong Wang. Gnfactor: Multi-task real robot learning with generalizable neural feature fields. InConferenceon robotlearning, pages 284–301. PMLR, 2023

    2023

  50. [50]

    3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations.Robotics: Scienceand SystemsXX, 2024

    Yanjie Ze, Gu Zhang, Kangning Zhang, Chenyuan Hu, Muhan Wang, and Huazhe Xu. 3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations.Robotics: Scienceand SystemsXX, 2024

    2024

  51. [51]

    Generalizable humanoid manipulation with 3d diffusion policies

    Yanjie Ze, Zixuan Chen, Wenhao Wang, Tianyi Chen, Xialin He, Ying Yuan, Xue Bin Peng, and Jiajun Wu. Generalizable humanoid manipulation with 3d diffusion policies. In2025 IEEE/RSJ International Conferenceon IntelligentRobotsandSystems(IROS), pages 2873–2880. IEEE, 2025

    2025

  52. [52]

    Chain-of-action: Trajectory autoregressive modeling for robotic manipulation.arXiv preprint arXiv:2506.09990, 2025

    Wenbo Zhang, Tianrun Hu, Yanyuan Qiao, Hanbo Zhang, Yuchu Qin, Yang Li, Jiajun Liu, Tao Kong, Lingqiao Liu, and Xiao Ma. Chain-of-action: Trajectory autoregressive modeling for robotic manipulation.arXiv preprint arXiv:2506.09990, 2025

    arXiv 2025

  53. [53]

    Effective tuning strategies for generalist robot manipulation policies

    WenboZhang,YangLi,YanyuanQiao,SiyuanHuang,JiajunLiu,FerasDayoub,XiaoMa,andLingqiaoLiu. Effective tuning strategies for generalist robot manipulation policies. In2025 IEEE International Conferenceon Roboticsand Automation(ICRA), pages 7255–7262. IEEE, 2025

    2025

  54. [54]

    same-mode

    Tony Z Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manipulation with low-cost hardware.arXivpreprintarXiv:2304.13705, 2023. 15 Appendix A Theoretical Analysis of Compounding Errors As introduced in the method section of the main paper, Segmented Trajectory Modeling (STM) mitigates compounding errors that arise in co...

    Pith/arXiv arXiv 2023