Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Fixed execution horizons are unreliable for action-chunking robots; a training-free rule that replans at motion valleys lifts success from 57.8% to 64.2% in simulation and from 50.7% to 70.4% on real robots.

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

2026-08-02 12:40 UTC pith:SKBGTECH

load-bearing objection PACE is a genuinely useful execution-time idea with a strong evaluation, but the headline gain over fixed horizons is partially confounded by per-task threshold calibration. the 3 major comments →

arxiv 2606.00537 v2 pith:SKBGTECH submitted 2026-05-30 cs.RO

PACE: Phase-Aware Chunk Execution for Robot Policies with Action Chunking

Junnan Nie , Jiayi Li , Jiachen Zhang , Junyi Lao , Chenghao Liu , Tianle Zhang , Liang Lin , Songfang Huang This is my paper
classification cs.RO
keywords action chunkingexecution horizontest-time adaptationreplanningrobot manipulationspeed profilephase transitionsimitation learning
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 targets a decision that action-chunking robot policies leave unspecified: after a policy predicts a sequence of future actions, how many should the robot execute before it takes a new observation. It argues that this execution horizon has a large, task-dependent, non-monotonic effect on success, so no single fixed horizon is a reliable default. The proposed answer, PACE, is a training-free test-time rule that reads the predicted chunk's speed profile, locates low-speed valleys (the deceleration regions where manipulation phases transition), and replans there instead of at uniform intervals. Across 50 simulated bimanual manipulation tasks, PACE lifts the average success rate from 57.8% (the best fixed horizon) to 64.2%; on real-robot trials the average success rate rises from 50.7% to 70.4%. The reason to care: the improvement comes purely from how predictions are executed, requiring no retraining, no new sensor inputs, and no access to the policy's internals.

Core claim

The central claim is that the execution horizon should be selected per chunk from the predicted chunk's own kinematic structure rather than fixed globally. PACE constructs a speed profile for each executed arm from the predicted actions, smooths it, detects prominent low-speed valleys as candidate phase boundaries, and executes the prefix up to the earliest accepted valley—or up to a maximum horizon if no valley passes the acceptance threshold. Manipulation trajectories are not kinematically uniform: they consist of coherent motion segments separated by deceleration events such as contact preparation, grasping, release, or alignment, and imitation-learned policies preserve this structure in

What carries the argument

The central object is the predicted action chunk's speed profile: PACE maps each arm's predicted actions to a one-dimensional speed sequence, smooths it to suppress spurious fluctuations, and treats prominent low-speed valleys as candidate replanning boundaries. Candidate valleys from all executed arms are pooled, those whose prominence exceeds a threshold calibrated once from training demonstrations are accepted, and the chunk is executed up to the earliest accepted valley. This mechanism carries the argument because it converts a content-blind constant horizon into a chunk-conditioned online choice, and the matched-horizon results show that timing—not average query frequency—is what improv

Load-bearing premise

The method assumes that every useful replanning boundary shows up as a low-speed valley in the predicted speed profile; if a task's key decision points occur during fast motion, or the policy's predictions are too noisy for valleys to be detected, PACE has no signal to act on.

What would settle it

Run PACE on a task whose demonstrations contain no deceleration before the critical decision—for example, a continuous sweeping or stirring motion where the decisive event occurs at peak speed—and compare it against a fixed horizon matched to PACE's average execution length. If PACE does not beat the matched fixed schedule in that regime, the claim that low-speed valleys mark useful replanning boundaries is falsified for fast-decision tasks.

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

If this is right

  • PACE can be added to any already-trained action-chunking policy with no retraining, no auxiliary heads, and no access to internal attention, confidence, or denoising signals.
  • Fixed execution horizons are not transferable: the best horizon differs by task and varies non-monotonically within a task, so deployment-time rules are necessary rather than a convenience.
  • The gains come from where replanning happens, not from querying more often: PACE outperforms a fixed schedule matched to the same average execution length on all six representative tasks compared.
  • Longer training prediction horizons improve the executed prefix even when the unexecuted suffix is discarded, so the benefits of chunk length and execution rule compound.
  • Real-robot deployment improves under the same rule on both bimanual and single-arm platforms, suggesting the mechanism is not specific to one robot morphology.

Where Pith is reading between the lines

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

  • Extension: because PACE only reads the predicted action sequence, it should transfer to any chunk-generating policy family—action-tokenized models, world-action models, or other vision-language-action architectures—that preserves low-speed transitions; the paper evaluates one policy family, so this remains to be tested.
  • Extension: the per-task threshold calibrated from training demonstrations might be replaceable by a global or few-shot-calibrated setting for tasks with similar contact structure, which would remove the remaining per-task deployment step.
  • Extension: a natural combination is to override the valley rule with an uncertainty signal—when the policy is unsure, replan even earlier than the next valley—which could address tasks whose critical decisions occur at high speed rather than at decelerations.
  • Extension: because PACE discards the chunk suffix, it reduces open-loop commitment; this may make rollouts more robust to sudden disturbances mid-chunk, a regime the paper's experiments do not directly stress.

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 / 4 minor

Summary. This paper studies the execution horizon in action-chunked robot policies and proposes PACE, a test-time rule that selects, for each predicted chunk, the number of actions to execute before re-querying the policy. PACE computes a speed profile from the predicted action chunk, smooths it, detects low-speed valleys, and executes up to the earliest valley whose prominence exceeds a task-level threshold calibrated from training demonstrations; if none is accepted it executes a fallback maximum horizon. The paper reports that on 50 RoboTwin2.0 tasks with a π0.5 policy, PACE improves the average success rate from 57.8% for the best fixed global horizon to 64.2%, and in real-robot experiments improves average score from 60.7 to 77.7 and success rate from 50.7% to 70.4%. A matched-horizon diagnostic on six tasks and a real-robot rollout analysis are used to argue that the gains come from adaptive replanning timing rather than average query frequency.

Significance. If the claimed results hold, PACE is a practically valuable, plug-and-play execution layer that requires no retraining and no access to policy internals. The evaluation is unusually extensive for a test-time method: 900 episodes per task-method pair, 45,000 episodes per method for the 50-task comparison, real-robot trials on ALOHA and Franka, ablations, hyperparameter sensitivity, and a matched-horizon diagnostic. The observation that fixed execution horizons are task-dependent and non-monotonic in success is well supported and is an important deployment consideration for action-chunking policies. The main weakness is that the primary comparison does not give fixed-horizon baselines the same per-task calibration budget as PACE, and the kinematic-valley assumption is not directly validated. The paper is transparent about its limitations, including the fact that experiments use a single VLA policy family.

major comments (3)
  1. [Sec. 4.1/Table 1; Sec. 3.3] The primary comparison is confounded by calibration budget. PACE uses a task-level acceptance threshold δ_T calibrated from training demonstrations, while the fixed baselines in Sec. 4.1 are global constants H∈{5,25,50} not tuned per task. Since Fig. 1 shows success is task-dependent and non-monotonic in H, a per-task scalar can shift the aggregate without any online valley selection. The phrase "without task-specific horizon sweeps" does not control for this. Please add a fixed-horizon baseline that chooses H per task from training demonstrations only, using the same data and selection budget as δ_T, and report the 50-task average. Table 4 does not resolve this: it covers six tasks, uses a seed-0 fixed sweep against three-seed PACE, and matches average executed horizon, not per-task calibration budget.
  2. [Sec. 7; Sec. 3.2] The central kinematic assumption is acknowledged but not directly tested. PACE assumes that useful replanning boundaries are expressed as low-speed valleys in the predicted chunk, and all experiments use a single VLA family (π0.5). Since valley detection is the entire signal, I recommend a direct validation that detected valleys align with phase transitions (e.g., using demonstration segment annotations, or comparing replanning at valleys against matched non-valley points). This would also clarify whether the gains reflect the proposed mechanism or a policy-specific artifact.
  3. [Sec. 4.4; Table 4] The matched-horizon diagnostic uses a seed-0 fixed-horizon sweep for the fixed baseline and a three-seed full protocol for PACE, as acknowledged in Appendix D. The main text should state this asymmetry when presenting the matched-horizon gains; otherwise the reported improvements may be partly an artifact of evaluation noise. A three-seed matched fixed baseline is needed before the conclusion that adaptive replanning timing alone drives the gains is secure.
minor comments (4)
  1. [Sec. 3.3] H_max is not assigned a value or included in the sensitivity study; the paper should specify how it is set (presumably L=50) and whether the results are sensitive to it.
  2. [Abstract; throughout] The term "training-free" is slightly misleading because δ_T is calibrated from training demonstrations. Consider "without additional training" or "without retraining".
  3. [Table 3] Only d_min and ρ are varied in the sensitivity analysis. The smoothing width, prominence computation, and H_max are fixed; a brief statement justifying these choices would be helpful.
  4. [Sec. 4.1] Despite the large number of episodes, no confidence intervals or standard errors are reported. Adding them would help gauge the reliability of the 6.4-point aggregate gain.

Circularity Check

0 steps flagged

No significant circularity: PACE's central claim is empirical and self-contained; the reported gains are measured, not derived from a chain that reduces to its own inputs.

full rationale

PACE is not a derivation chain that maps inputs to predicted outcomes; it is a test-time heuristic evaluated by measured success rates. The only fitted quantity, δ_T, is calibrated from training demonstrations (Sec. 3.3), and the paper explicitly states that this calibration does not use evaluation rollouts, success labels, or fixed-horizon sweeps. The headline gain (57.8% to 64.2%) is a measured outcome, so it cannot be statistically forced by the calibration. The central kinematic assumption—that useful replanning boundaries are reflected as low-speed valleys—is a stated premise, not a conclusion derived from PACE, and the paper itself flags it as a limitation in Sec. 7. References to RoboTwin2.0, π0.5, and prior phase-segmentation work are external benchmarks or prior literature, not self-citation chains that supply the result. The per-task δ_T versus global fixed-H comparison is a legitimate experimental-fairness concern, but an unfair comparison is not circularity: no equation or fitted parameter is renamed as the output. Therefore no step meets the required standard of exhibiting an identity between an output and an input.

Axiom & Free-Parameter Ledger

5 free parameters · 3 axioms · 0 invented entities

PACE introduces no new physical entities. It relies on calibrated thresholds and assumptions about kinematic structure. The free parameters are inherent to the method but are mostly keyed to the training set, which is legitimate but introduces a per-task tuning advantage.

free parameters (5)
  • δ_T (acceptance threshold) = not reported
    Calibrated per task from training demonstrations; controls which valleys are accepted as boundaries (Sec 3.3).
  • d_min (minimum valley separation) = 10 (default)
    Hyperparameter in valley detection; varied in sensitivity analysis (Table 3).
  • ρ (calibration percentile) = 5 (default)
    Determines the strictness of δ_T; varied in sensitivity analysis (Table 3).
  • H_max (maximum execution horizon) = L=50
    Maximum allowed execution horizon; set to the prediction horizon.
  • Smoothing operator parameters = not specified
    The smoothing operator S is not defined; its parameters are implicit and likely chosen by the authors.
axioms (3)
  • domain assumption Predicted action chunks preserve the phase structure of expert demonstrations
    PACE relies on low-speed valleys in predicted chunks signaling phase transitions (Sec 3.2). This is not proven but argued from imitation learning.
  • domain assumption Low-speed valleys are reliable replanning points
    The method assumes that executing up to a valley improves performance over fixed horizons (Sec 3.3).
  • domain assumption The speed profile is a sufficient statistic for phase boundaries
    PACE uses only kinematic speed, ignoring forces, images, or other signals (Sec 3.2).

pith-pipeline@v1.3.0-alltime-deepseek · 205 in / 7041 out tokens · 104600 ms · 2026-08-02T12:40:02.658178+00:00 · methodology

0 comments
read the original abstract

Recent vision-language-action and diffusion-based robot policies often use action chunking, where each policy query predicts a sequence of future actions and the robot executes an open-loop prefix before re-querying. While this interface improves local motion continuity, deployment still requires choosing the execution horizon: how much of each predicted chunk should be executed before acquiring a new observation. However, our experiments show that success is strongly task-dependent and non-monotonic with respect to the execution horizon, making a single constant horizon an unreliable deployment rule. We propose PACE (Phase-Aware Chunk Execution), a training-free test-time execution method that selects the execution horizon online from the predicted chunk itself. PACE exploits the phase-dependent kinematic structure of manipulation trajectories by identifying low-speed transition points in the predicted speed profile and using them as candidate replanning boundaries. Because PACE uses only the predicted action chunk, it is plug-and-play and requires no retraining or access to policy internals. We validate PACE through large-scale evaluations in both simulation and real-robot settings. On 50 RoboTwin2.0 tasks, PACE raises the average success rate from 57.8% to 64.2%. In real-robot experiments on bimanual ALOHA and single-arm Franka platforms, PACE improves the average task score from 60.7 to 77.7 and the average success rate from 50.7% to 70.4%. Ablations and rollout-level analyses show that PACE adapts execution horizons across manipulation phases, shortening near transitions while preserving longer execution during coherent motion.

Figures

Figures reproduced from arXiv: 2606.00537 by Chenghao Liu, Jiachen Zhang, Jiayi Li, Junnan Nie, Junyi Lao, Liang Lin, Songfang Huang, Tianle Zhang.

Figure 1
Figure 1. Figure 1: Overview of PACE. Fixed horizons can be unreliable because success varies with H across tasks. PACE selects the executed prefix online from each predicted chunk, improving performance in simulation and real-robot experiments while adapting the horizon in a rollout. †Corresponding author. Emails: jnnie25@stu.pku.edu.cn, 23121254@bjtu.edu.cn, z89498323286@gmail.com, jylao25@stu.pku.edu.cn, chliu@stu.pku.edu.… view at source ↗
Figure 2
Figure 2. Figure 2: PACE framework. PACE selects an execution hori￾zon from low-speed valleys in the predicted chunk’s smoothed speed profile. The robot executes the selected prefix, discards the suffix, and then queries the policy again. chunk execution [2, 23]. Another line modifies training objec￾tives or model design so that policies can reason over multiple horizons [11]. Closest to our setting, AutoHorizon selects execu… view at source ↗
Figure 3
Figure 3. Figure 3: Rollout-level behavior of PACE. The rollout is from place_shoe. Top: head and front camera observations at the six replanning timesteps. Middle: selected execution horizons between consecutive queries. Bottom: predicted action chunks along the rollout timeline, where solid segments are executed prefixes and dashed segments are discarded suffixes. Vertical dashed lines mark replanning boundaries. substantia… view at source ↗
Figure 4
Figure 4. Figure 4: PACE compared with fixed-horizon sweeps. Green curves show the seed-0 diagnostic sweep of fixed-horizon execution as H is varied from 1 to 50 on each task. The red star marks PACE under the full three-seed evaluation: its horizontal coordinate is the mean executed horizon averaged over policy queries, and its vertical coordinate is the PACE success rate. PACE is shown as a point only for visualization and … view at source ↗
Figure 5
Figure 5. Figure 5: Successful real-robot rollout on stack_bowls. Blue markers indicate policy-query timesteps, and green labels indicate the execution horizon selected between consecutive queries. PACE selects long horizons during approach and transport, shortens the horizon near contact-sensitive stacking alignment, and expands it again once a coherent motion segment becomes available [PITH_FULL_IMAGE:figures/full_fig_p008… view at source ↗
Figure 6
Figure 6. Figure 6: Failure case on put_pen_into_pencil_case. The rollout is from the ALOHA robot. Blue markers indicate policy-query timesteps, and green labels indicate selected horizon lengths. The pencil case remains only partially opened while the right arm moves the pen toward it, showing a failure of the base policy rather than feedback timing. 5.3 Failure Case and Scope of Test-Time Execution Control [PITH_FULL_IMAGE… view at source ↗
Figure 7
Figure 7. Figure 7: Expanded training prediction horizon ablation. Each cell shows the success-rate gain of training horizon Htrain (columns) relative to the shortest feasible training horizon for a given evaluation horizon Heval (rows), i.e., Htrain = Heval. Blue indicates a gain, orange a loss; the diagonal is zero by construction. Cells below the diagonal are infeasible and left blank. episodes. Repeating the full sweep fo… view at source ↗
Figure 8
Figure 8. Figure 8: Initial frames of the RoboChallenge tasks. Left: put_pen_into_pencil_case. Right: stack_bowls. Both tasks are evaluated on an ALOHA robot using the same fine-tuned checkpoint within each task; only the test-time execution rule differs between the baseline and PACE. For place_object_on_plate, the Franka robot must pick up a specified object and place it fully inside a target plate. We use five object varian… view at source ↗
Figure 9
Figure 9. Figure 9: In-lab place_object_on_plate setup. The task uses five objects—corn, cabbage, green pepper, red pepper, and garlic—and a fixed target plate. A single fine-tuned π0.5 checkpoint is used across all object variants, and each method is evaluated over 5 × 20 real-robot trials. put_pen_into_pencil_case. The task is to place the pen into the pencil case. In our evaluation setup and rollout videos, the left grippe… view at source ↗
Figure 10
Figure 10. Figure 10: PACE rollout visualization on place_can_basket. The selected execution horizon adapts to the local phase structure of the rollout, with shorter prefixes near the placement transition and longer prefixes during smooth motion segments. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: PACE rollout visualization on hanging_mug. The execution horizon is shortened around the task transition that requires precise alignment and interaction, and lengthened during more stable motion segments. scan_object. This rollout highlights a task with a more gradual motion structure. PACE continues to adapt the execution horizon online, using longer prefixes during steady motion and shorter prefixes whe… view at source ↗
Figure 12
Figure 12. Figure 12: PACE rollout visualization on scan_object. PACE selects execution horizons from the predicted kinematic profile, allowing the rollout to keep longer open-loop segments when motion is smooth and to replan earlier around phase transitions. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: PACE rollout visualization on RoboChallenge stack_bowls. The selected execution horizon is longer during smooth approach and transport, and shorter near the stacking phase, reflecting the same phase-aware replanning rule used throughout the paper. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_13.png] view at source ↗

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. TRACT: Temporally Routed Action Chunks with Chronological Phase Authority for Contact-Rich Manipulation

    cs.RO 2026-07 conditional novelty 6.0

    Phase-aware action chunks with an intra-chunk boundary ('routing') plus a causal response-deficit integrator raised real-robot wiping success on one fixed task from 3/10 to 10/10 across ten trials per variant.

Reference graph

Works this paper leans on

28 extracted references · 20 linked inside Pith · cited by 1 Pith paper

  1. [1]

    π0: A Vision-Language-Action Flow Model for General Robot Control

    Kevin Black et al. π0: A Vision-Language-Action Flow Model for General Robot Control. 2024.doi: 10.48550/ arXiv.2410.24164

    Pith/arXiv arXiv doi:10.48550/arxiv.2410.24164 2024

  2. [3]

    2022.doi: 10

    Anthony Brohan et al.RT-1: Robotics Transformer for Real-World Control at Scale. 2022.doi: 10 . 48550 / arXiv.2212.06817

    Pith/arXiv arXiv doi:10.48550/arxiv.2212.06817 2022

  3. [4]

    2023.doi:10.48550/arXiv.2307.15818

    Anthony Brohan et al.RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control. 2023.doi:10.48550/arXiv.2307.15818

    Pith/arXiv arXiv doi:10.48550/arxiv.2307.15818 2023

  4. [5]

    2025.doi: 10.48550/arXiv.2506

    Jun Cen et al.WorldVLA: Towards Autoregressive Ac- tion World Model. 2025.doi: 10.48550/arXiv.2506. 21539

    doi:10.48550/arxiv.2506 2025

  5. [6]

    Tianxing Chen et al.RoboTwin 2.0: A Scalable Data Generator and Benchmark with Strong Domain Random- ization for Robust Bimanual Robotic Manipulation. 2025. doi:10.48550/arXiv.2506.18088

    Pith/arXiv arXiv doi:10.48550/arxiv.2506.18088 2025

  6. [7]

    2023.doi: 10

    Cheng Chi et al.Diffusion Policy: Visuomotor Policy Learning via Action Diffusion. 2023.doi: 10 . 48550 / arXiv.2303.04137

    Pith/arXiv arXiv doi:10.48550/arxiv.2303.04137 2023

  7. [8]

    2023.doi: 10.48550/arXiv.2303

    Danny Driess et al.PaLM-E: An Embodied Multimodal Language Model. 2023.doi: 10.48550/arXiv.2303. 03378

    doi:10.48550/arxiv.2303 2023

  8. [9]

    2024.doi:10.48550/arXiv.2401.02117

    Zipeng Fu et al.Mobile ALOHA: Learning Bimanual Mobile Manipulation with Low-Cost Whole-Body Teleop- eration. 2024.doi:10.48550/arXiv.2401.02117

    Pith/arXiv arXiv doi:10.48550/arxiv.2401.02117 2024

  9. [10]

    π0.5: A Vision-Language- Action Model with Open-World Generalization

    Physical Intelligence et al. π0.5: A Vision-Language- Action Model with Open-World Generalization. 2025.doi: 10.48550/arXiv.2504.16054

    Pith/arXiv arXiv doi:10.48550/arxiv.2504.16054 2025

  10. [11]

    2025.doi:10.48550/arXiv.2511.19433

    Dong Jing et al.Mixture of Horizons in Action Chunking. 2025.doi:10.48550/arXiv.2511.19433

    Pith/arXiv arXiv doi:10.48550/arxiv.2511.19433 2025

  11. [12]

    2024.doi: 10

    Alexander Khazatsky et al.DROID: A Large-Scale In- The-Wild Robot Manipulation Dataset. 2024.doi: 10 . 48550/arXiv.2403.12945

    Pith/arXiv arXiv doi:10.48550/arxiv.2403.12945 2024

  12. [13]

    2025.doi: 10

    Moo Jin Kim et al.Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success. 2025.doi: 10. 48550/arXiv.2502.19645

    Pith/arXiv arXiv doi:10.48550/arxiv.2502.19645 2025

  13. [14]

    2024.doi: 10.48550/arXiv

    Moo Jin Kim et al.OpenVLA: An Open-Source Vision- Language-Action Model. 2024.doi: 10.48550/arXiv. 2406.09246

    Pith/arXiv arXiv doi:10.48550/arxiv.2406.09246 2024

  14. [15]

    Towards Learning Hierarchical Skills for Multi-Phase Manipulation Tasks

    Oliver Kroemer et al. “Towards Learning Hierarchical Skills for Multi-Phase Manipulation Tasks”. In:2015 IEEE International Conference on Robotics and Automa- tion (ICRA). 2015.doi: 10.1109/ICRA.2015.7139389

    arXiv 2015

  15. [16]

    Autonomous Framework for Segmenting Robot Trajectories of Ma- nipulation Task

    Sang Hyoung Lee, Il Hong Suh, et al. “Autonomous Framework for Segmenting Robot Trajectories of Ma- nipulation Task”. In: (2015).doi: 10 . 1007 / s10514 - 014-9397-9

    2015

  16. [17]

    2024.doi: 10.48550/ arXiv.2410.07864

    Songming Liu et al.RDT-1B: A Diffusion Foundation Model for Bimanual Manipulation. 2024.doi: 10.48550/ arXiv.2410.07864

    Pith/arXiv arXiv doi:10.48550/arxiv.2410.07864 2024

  17. [18]

    2023.doi:10.48550/arXiv.2310.08864

    Open X-Embodiment Collaboration et al.Open X- Embodiment: Robotic Learning Datasets and RT-X Mod- els. 2023.doi:10.48550/arXiv.2310.08864

    Pith/arXiv arXiv doi:10.48550/arxiv.2310.08864 2023

  18. [19]

    2025.doi: 10.48550/ arXiv.2501.09747

    Karl Pertsch et al.FAST: Efficient Action Tokenization for Vision-Language-Action Models. 2025.doi: 10.48550/ arXiv.2501.09747

    Pith/arXiv arXiv doi:10.48550/arxiv.2501.09747 2025

  19. [20]

    Mustafa Shukor et al.SmolVLA: A Vision-Language- Action Model for Affordable and Efficient Robotics. 2025. doi:10.48550/arXiv.2506.01844

    Pith/arXiv arXiv doi:10.48550/arxiv.2506.01844 2025

  20. [21]

    2024.doi: 10.48550/arXiv.2405

    Octo Model Team et al.Octo: An Open-Source Gener- alist Robot Policy. 2024.doi: 10.48550/arXiv.2405. 12213

    doi:10.48550/arxiv.2405 2024

  21. [22]

    2023.doi: 10.48550/arXiv.2308

    Homer Walke et al.BridgeData V2: A Dataset for Robot Learning at Scale. 2023.doi: 10.48550/arXiv.2308. 12952

    doi:10.48550/arxiv.2308 2023

  22. [23]

    2026.doi: 10.48550/arXiv

    Haoxuan Wang et al.Real-Time Robot Execution with Masked Action Chunking. 2026.doi: 10.48550/arXiv. 2601.20130

    doi:10.48550/arxiv.2601.20130 2026

  23. [24]

    2026.doi: 10.48550/arXiv.2602.21445

    Haoxuan Wang et al.VLA Knows Its Limits. 2026.doi: 10.48550/arXiv.2602.21445

    Pith/arXiv arXiv doi:10.48550/arxiv.2602.21445 2026

  24. [25]

    2025.doi: 10

    Adina Yakefu et al.RoboChallenge: Large-scale Real- robot Evaluation of Embodied Policies. 2025.doi: 10. 48550/arXiv.2510.17950

    doi:10.48550/arxiv.2510.17950 2025

  25. [26]

    2026.doi:10.48550/arXiv.2602.15922

    Seonghyeon Ye et al.World Action Models are Zero-shot Policies. 2026.doi:10.48550/arXiv.2602.15922

    Pith/arXiv arXiv doi:10.48550/arxiv.2602.15922 2026

  26. [27]

    2024.doi:10.48550/arXiv.2403.03954

    Yanjie Ze et al.3D Diffusion Policy: Generalizable Vi- suomotor Policy Learning via Simple 3D Representations. 2024.doi:10.48550/arXiv.2403.03954

    Pith/arXiv arXiv doi:10.48550/arxiv.2403.03954 2024

  27. [28]

    2026.doi: 10

    Tianle Zhang et al.JoyAI-RA 0.1: A Foundation Model for Robotic Autonomy. 2026.doi: 10 . 48550 / arXiv . 2604.20100

    Pith/arXiv arXiv doi:10.48550/arxiv.2604.20100 2026

  28. [29]

    Zhao et al.Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware

    Tony Z. Zhao et al.Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware. 2023.doi: 10. 48550/arXiv.2304.13705. 10 A PACE ImplementationNotes PACE is applied only at test time and does not modify or retrain the base policy. After each policy query, it analyzes the predicted action chunk and selects how many actions to execute before the next ...

    Pith/arXiv arXiv doi:10.48550/arxiv.2304.13705 2023