PaperlayerView exact revision on arXiv

Source-linked AI summary

ChunkTrust: Adapting Execution Horizons for Robot Policies with Action-Expert Evidence

Fanding Huang, Jingyan Jiang, Shifeng Bao, Mingkang Pu, Shiwei Li, Jing Xu, Shijia Xu, Guanbo Huang, Chenghao Gu, Yuzhi Huang, Chenxin Li, Faisal Nadeem Khan, Huan Yang, Yan Wang, Cheng Chi, Zhi Wang

arXiv:2609.39754v1cs.RO

TL;DR

Robot policies predict action chunks, but the appropriate execution horizon changes with task phase, and existing evidence may not capture compatibility with motion already executed. ChunkTrust combines generation stability and executed-motion continuity with online memory and an optional learned prior. AHS improves aggregate success across evaluated simulation configurations and improves the real-robot process score in the reported tasks.

  • Problem

    Execution horizons must adapt as task phase changes, while existing approaches may miss whether an internally consistent prediction fits motion already executed.

    Source: §1 · p. 2

  • Method

    ChunkTrust combines spectral stability and inter-chunk continuity in AHS with episode-local Beta memory, while optional QHA learns a context-conditioned prior without changing the base policy.

    Source: §1 · pp. 2–3

  • Results

    AHS improves aggregate success across evaluated simulation base-policy configurations, including +6.80 percentage points on the complete 50-task RoboTwin2.0 suite and +9.67 percentage points with Qwen3GR00T on RoboCasa.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The reported evaluations support action-expert evidence for adaptive horizons across simulation benchmarks and real-world manipulation tasks.

    Source: §1 · p. 2, §6 · p. 10

    AHS requires access to action chunks and generation-time velocity traces, and the evaluation does not cover all embodiments, safety-critical tasks, or long-horizon mobile manipulation.

    Source: G DISCUSSION AND LIMITATIONS · p. 34

Abstract

from arXiv · show

Robot foundation policies predict action chunks, but how many actions to execute before replanning depends on the current task phase. We introduce ChunkTrust, which treats the execution horizon as a latent variable inferred from action-expert evidence rather than a fixed hyperparameter. Its training-free Action-aware Horizon Selector (AHS) combines intra-chunk spectral stability of generation traces with inter-chunk continuity between executed history and predicted actions. An online Beta posterior with kernel forgetting tracks horizon preferences across replans. A lightweight Query-based Horizon Adapter (QHA) optionally learns a context-conditioned dense prior from complementary evidence, fused with current evidence and episode-local Beta memory while the base policy remains frozen. Across RoboTwin2.0 and RoboCasa GR1 Tabletop, AHS improves overall task-averaged success for each evaluated base-policy configuration, including gains of +6.80 percentage points on $\pi_{0.5}$ over all 50 RoboTwin2.0 tasks and +9.67 percentage points on Qwen3GR00T in RoboCasa. AHS+QHA raises the gain over Base to +9.44 percentage points on the eight-task $\pi_{0.5}$ evaluation. On four real-world household tasks, AHS improves the equal-task mean normalized process score from 50.4% to 57.5%. Ablations examine the contributions of both evidence terms, temporal memory, and the learned prior. Project page is https://hf618.github.io/ChunkTrust.github.io/

1 INTRODUCTION

ChunkTrust addresses the need to choose how many predicted actions to execute before replanning as task phases change. It combines complementary action-expert evidence with online memory and an optional learned horizon prior, then evaluates the approach in simulation and on real robots.

1 INTRODUCTION

  • The execution horizon must balance smooth progress from longer open-loop execution against delayed correction, while shorter execution increases feedback but can disrupt coherent motion.The preferred horizon changes by task phase: free-space approach may tolerate longer execution, while contact or delicate transport calls for earlier replanning.Source: §1 · p. 2
  • Existing adaptive approaches use signals such as attention, entropy, cross-horizon agreement, denoising trajectories, or learned replanning decisions, but internally consistent predictions may still conflict with executed motion.The paper identifies noisy replan-level evidence and recurring execution contexts as motivation for complementary evidence and reusable knowledge.Source: §1 · p. 2
  • ChunkTrust pairs spectral stability during generation with continuity between predicted actions and recently executed motion.Retrospective analysis found higher median spectral instability and boundary variation in failed episodes, with the highest failure rate when both risks were elevated.Source: §1 · p. 2
  • AHS accumulates evidence in an episode-local Beta state with forgetting, while QHA learns reusable context-conditioned horizon preferences without changing the base policy.At deployment, QHA supplies a dense prior that complements online AHS evidence.Source: §1 · pp. 2–3
  • +6.80 percentage points on the complete 50-task RoboTwin2.0 suite and +9.67 percentage points with Qwen3GR00T on RoboCasa are AHS gains in aggregate success.AHS+QHA further improves aggregate success on the eight-task evaluation; on real robots, AHS raises equal-task mean normalized process score by 7.1 percentage points.Source: §1 · p. 2

2 RELATED WORK

Related work adapts robot execution through reactive generation, evidence-based horizon selection, and learned horizon policies. ChunkTrust instead selects a prefix at replanning time from action-expert evidence and executed history, then combines a learned prior with current evidence and episode-local memory.

Action generation and reactive execution

  • ChunkTrust selects a prefix at each replan using action-expert evidence and executed history, without modifying generated actions or monitoring new observations during that prefix.This contrasts with approaches that alter action generation or use execution monitors to react to scene changes or drift.Source: §2 · p. 3

Evidence-based horizon selection

  • ChunkTrust pairs spectral variation during generation with speed variation after stitching a candidate prefix to executed history, and AHS accumulates both signals in horizon-indexed Beta states with forgetting.Related methods use coherence, cross-horizon consensus, action entropy, denoising variation, geometric profiles, or low-speed transitions as horizon evidence.Source: §2 · p. 3

Learned horizon selection

  • Unlike learned horizon approaches based on reinforcement learning, joint latent and chunk selection, or observation-sensitivity forecasts, QHA learns a context-conditioned prior from complementary action-expert evidence.At deployment, QHA combines that prior with current evidence and episode-local Beta memory.Source: §2 · p. 3

3 HORIZON-AWARE EVIDENCE FROM THE ACTION EXPERT

The section defines candidate execution prefixes and develops two complementary signals: generation-time spectral stability and continuity with executed motion. Their fused evidence is associated with episode failure risk and supplies the basis for horizon inference.

3.1 PRELIMINARIES

  • At each replan, a frozen policy receives visual observation, proprioceptive state, and instruction, and the controller scores candidate prefixes using generation stability and compatibility with executed history.Expected-round selection can return intermediate integer lengths rather than only grid points.Source: §3.1 · pp. 3–4

3.2 INTRA-CHUNK SPECTRAL STABILITY

  • Intra-chunk instability measures RMS deviation of high-frequency velocity-prefix spectrum across denoising steps from an early baseline, with lower deviation receiving higher normalized quality.The Fourier transform is applied along the action horizon; the early window is a reference and remains included in the RMS calculation.Source: §3.2 · pp. 4–5

3.3 INTER-CHUNK CONTINUITY

  • Inter-chunk discontinuity is the speed coefficient of variation along a stitched history-and-candidate trajectory; higher variation can reflect boundary jumps or stop-and-go motion.A smaller value indicates a more uniform stitched trajectory, and the score is normalized among candidates at the current replan.Source: §3.3 · p. 5

3.4 DUAL-EVIDENCE FUSION

  • Dual-evidence fusion combines internal stability and compatibility with recent motion using qmix,t(k) = (1 − λ)qintra,t(k) + λqinter,t(k), with λ = 0.5 as default.The combined score is evidence for horizon inference, not a deterministic execution rule or calibrated task-success probability.Source: §3.4 · p. 5
  • Failure rates rise from 50.2% when both evidence risks are low to 97.4% when both are high in the fixed-horizon diagnostic episodes.The comparison uses 1,600 RoboTwin2.0 episodes with π0 and fixed K = H = 50; the association does not show that a low-risk prefix guarantees success.Source: Figure 2 · p. 4, §3.4 · p. 5

4 ACTION-AWARE HORIZON SELECTION AND QUERY-BASED ADAPTATION

AHS selects and executes action prefixes using current evidence and episode-local reliability memory that can adapt as task phases change. QHA learns context-conditioned horizon preferences across episodes and combines them with online evidence at deployment.

4.1 ACTION-AWARE HORIZON SELECTOR

  • AHS combines current candidate quality with sampled episode-local Beta reliability to select an action prefix, then executes it and replans from the next observation.If scores degenerate, it falls back to a uniform distribution over valid candidates; clipping keeps the prefix within the available prediction.Source: §4.1 · p. 6
  • Kernel sharing spreads soft feedback across nearby horizons, while exponential forgetting discounts earlier evidence so AHS can adapt as task phases change.The feedback is derived from action-expert quality, not an observed task-success label.Source: §4.1 · p. 6

4.2 TRAINING-TIME QUERY-BASED HORIZON ADAPTER

  • QHA learns context-conditioned horizon preferences from VLM context and action-latent tokens, encoding preferences across training episodes as cross-episode memory.Its learned horizon queries use bridge cross-attention over both token streams; only QHA is trained.Source: §4.2 · p. 6

4 ACTION-AWARE HORIZON SELECTION AND QUERY-BASED ADAPTATION

  • Table 1 reports simulation success rates for AHS versus Base using the same policy checkpoint, with Δ defined as AHS success rate minus Base success rate.The caption identifies benchmark-level performance but does not provide table cell values.Source: Table 1 · p. 7

4.2 TRAINING-TIME QUERY-BASED HORIZON ADAPTER

  • At deployment, QHA supplies a dense prior combined with dense Beta reliability evidence, while online traces and AHS candidate evaluation remain necessary for correction.The prior provides a context-conditioned preference before episode-local evidence has accumulated much.Source: §4.2 · pp. 6–7

5 EXPERIMENTS

AHS improves aggregate success across the evaluated simulation policy configurations and improves real-world process scores; QHA adds gains on training-task and held-out evaluations, with task-level regressions and costs also examined.

5 EXPERIMENTS

  • AHS improves aggregate success across task-specific, multitask, and zero-shot benchmark evaluation regimes, while those regimes are distinct checkpoint cohorts rather than matched-policy comparisons.The Base–AHS comparisons hold the policy checkpoint fixed; AHS changes execution without updating policy weights.Source: §5.1 · p. 7, §5 · p. 7

5.1 AHS ACROSS SIMULATION BENCHMARKS

  • AHS raises full-suite RoboTwin2.0 π0.5 success from 56.70% to 63.50% (+6.80 percentage points), and improves aggregate success across evaluated policies on both simulation benchmarks.On RoboCasa, Qwen3GR00T rises from 47.83% to 57.50%; the eight-task RoboTwin2.0 evaluation also improves for π0, π0.5, and Fast-WAM.Source: §5.1 · p. 7

5.2 QHA AUGMENTATION AND TRANSFER

  • AHS+QHA raises π0.5 overall success from 36.25% to 39.06% on QHA training tasks and from 40.00% to 42.75% on two held-out tasks.The training-task evaluation also improves π0 from 21.38% to 24.88%; on held-out tasks, Blocks Ranking RGB rises from 38.50% to 43.50%, while Place Bread Basket rises from 41.50% to 42.00%.Source: §5.2 · p. 8

5.3 REAL-WORLD DEPLOYMENT

  • On four real-world bimanual tasks, AHS raises the equal-task mean normalized process score from 50.4% to 57.5%.Scores average predefined sub-steps and rollouts rather than measuring binary success; Duck Toy to Drawer has the largest gain, from 50.0% to 63.3%.Source: §5.3 · p. 8, Figure 4 · p. 9

5.4 ABLATION STUDIES

  • With five candidates, π0.5 success rises from 22.0% at fixed K = 50 to 33.0%, versus 34.9% with 50 candidates; added per-replan cost increases from 0.67 to 3.45 ms.The sparse candidate set captures most of the gain, and the reported episode times cover successes only, so they do not establish all-attempt cost.Source: §5.4 · pp. 8–9, Table 3 · p. 9
  • On π0, full AHS reaches 15.0% overall and the best Hard result, 5.75%; adding QHA raises overall success to 15.8% while Hard falls to 4.00%.Combining both evidence terms without memory gives 14.6% overall and 3.75% on Hard.Source: §5.4 · p. 9, Table 4 · p. 9
  • With RTC at +200 ms, AHS waiting falls to 0.344 s per episode and observation age declines, while success rises on both settings; RTC does not uniformly improve success over synchronous AHS.On Easy, RTC success is 42.75% versus 44.75% for synchronous AHS, and the asynchronous setup requires more policy calls and model computation.Source: §5.4 · p. 9, Figure 5 · p. 10

6 CONCLUSION

ChunkTrust adapts execution horizons from action-expert evidence using online memory and an optional learned prior, without changing base-policy weights. The evaluation spans simulation benchmarks and real-world manipulation, while its scope includes accessible generation traces and nonuniform task-level gains.

6 CONCLUSION

  • ChunkTrust combines spectral stability, inter-chunk continuity, episode-local temporal memory, and an optional context-conditioned horizon prior while leaving the base policy unchanged.The method requires accessible generation traces; gains are not uniform across tasks, and broader real-world validation across hardware and domain shifts remains future work.Source: §6 · p. 10

A.1 REAL-WORLD EXPERIMENTS

  • Real-world evaluation uses an AgileX COBOT Magic configured as an ALOHA-style bimanual system, with four 6-DoF arms and three RealSense D435 cameras.Two leader arms support teleoperation and two follower arms support data collection and autonomous execution.Source: §A.1 · p. 17, Figure 6 · p. 18
  • The four real-world tasks are evaluated with π0.5 using fixed K = 25 or AHS over 15 rollouts per task, with object positions varied across rollouts.Process scores average sub-step ratings and rollouts; each sub-step receives 0 for failure, 0.5 for recovered or imperfect completion, and 1 for smooth, accurate completion.Source: §A.1 · p. 17

A.2 SIMULATION EXPERIMENTS

  • RoboTwin2.0 evaluation covers a 50-task cohort and a separate eight-task cohort, with Easy clean scenes and Hard randomized scenes; RoboCasa evaluation covers all 24 tasks.Success rates are averaged equally over the tasks and settings in each cohort; the π0.5 RoboCasa control uses 50 rollouts per task.Source: §A.2 · p. 17

A.3 BASE POLICY CHECKPOINTS

  • Checkpoint provenance varies across evaluations: policies include task-specific fine-tuning, multitask post-training, and released checkpoints, so absolute scores across regimes are not matched-policy comparisons.Within each benchmark, Base and AHS use identical policy weights; the two multitask controls share an architecture, not weights across benchmarks.Source: §A.3 · pp. 18–19

A.4 AHS HYPERPARAMETER CONFIGURATIONS

  • Candidate horizons and continuity windows vary by policy and benchmark, including {10, 20, 30, 40, 50} for RoboTwin2.0 π0 and π0.5 and {4, 8, 12, 16} for RoboCasa.Without sufficient executed history, AHS uses intra-chunk-only scoring.Source: §A.4 · p. 19

A.5 EVALUATION PROTOCOL

  • AHS resets its online posterior to the prior at each episode, while learned QHA weights are reused across episodes without online training.Reported success rates are raw rollout success percentages and equal-weight averages over stated task and setting groups; absolute differences are percentage points.Source: §A.5 · p. 19
  • Only 62 of 100 task–setting conditions in the 50-task evaluation use exact same-seed and instruction pairing between Base and AHS; the rest use deterministic fallback identities.Fallback identities are selected without outcome information after repeated scene-construction failures, and identical episode identities are not assumed for that group.Source: §A.5 · p. 19

A.6 QHA TRAINING CONFIGURATIONS

QHA learns a context-conditioned horizon prior from evidence generated by frozen base policies and combines it with online evidence at deployment. The appendix also describes the training and evaluation setup for evidence distributions and boundary profiles.

A.6 QHA TRAINING CONFIGURATIONS

  • QHA is trained on eight tasks per policy family with task-specific base generators frozen, using balanced batches and context, robot-state, action-window, and action-latent inputs.Its teacher is computed over horizons 1–50 from current evidence without episode-local Beta memory.Source: §A.6 · pp. 19–20
  • The QHA bridge updates eight learned queries through masked context and action-token attention, then query pooling and an MLP produce 50 horizon logits and a softmax.Optimization updates only QHA parameters; the base flow-matching loss is zero.Source: §A.6 · p. 20
  • At deployment, the eight-task QHA uses a learned prior with sparse online evidence interpolated across horizons, while the held-out evaluation computes dense evidence over horizons 1–50.Both configurations maintain dense Beta states and use temperature scaling, expected-round selection, and uniform exploration.Source: §A.6 · p. 20

B.1 EVIDENCE DISTRIBUTIONS AND BOUNDARY PROFILES

  • The evidence-distribution analysis scores five candidate horizons on the same 1,600 recorded π0 episodes, all executed at fixed K = H = 50 rather than as separate horizon trials.The episodes comprise eight tasks, two settings, and 100 episodes per task-setting pair, with 291 successes and 1,309 failures.Source: §B.1 · p. 21
  • For each candidate, finite scores are averaged over episode replans and then summarized by episode-level medians and IQRs; valid counts can differ because inter-chunk evidence needs executed history.The intra score uses zero-padded prefixes, a three-step baseline, and RMS over ten denoising steps.Source: §B.1 · p. 21
  • Boundary profiles use normalized first-difference speeds in a 60-step stitched history window, averaged within episodes before cross-episode summaries.This aggregation prevents longer episodes from receiving extra weight.Source: §B.1 · p. 21

B.2 JOINT RISK AND EPISODE OUTCOMES

This analysis defines episode-level intra- and inter-risk groups at K = 50 and compares their outcomes. The reported patterns are pooled retrospective associations, not causal or calibrated predictions.

B.2 JOINT RISK AND EPISODE OUTCOMES

  • At K = 50, each episode’s intra/inter risk is the 75th percentile of 1 − qintra,t(50) or 1 − qinter,t(50) across valid replans.Values below the pooled median thresholds (0.98168436 and 0.96278395) are low; ties are high, so group sizes can differ.Source: §B.2 · p. 21
  • The outcome map smooths success and failure counts separately on a padded 23 × 23 grid before division.Cells below 10^-3 occupancy are hidden.Source: §B.2 · p. 21
  • The pooled retrospective associations do not establish causality, calibrated failure prediction, within-task effects independent of difficulty, or current-replan prefix reliability.Adaptive execution benefits are evaluated separately in policy comparisons and ablations, not inferred from this heatmap.Source: §B.2 · p. 21
  • Table 6 reproduces Figure 2(c)’s exact median-split group counts and failure rates.The supporting passage does not supply those cell values.Source: Table 6 · p. 22

C.1 QHA TRANSFER TO HELD-OUT TASKS

These evaluations report QHA transfer, full-suite and policy-family AHS comparisons, RoboCasa results, and compact selector controls. Cohort and pairing boundaries qualify what each comparison establishes.

C.1 QHA TRANSFER TO HELD-OUT TASKS

  • QHA and AHS choices differ at 77.02% of 12,573 same-state replans, with a mean absolute gap of 1.83 action steps; this does not establish complementarity or explain gains.Easy and Hard disagreement rates are 77.17% and 76.90%, respectively.Source: §C.1 · p. 22

C.2 COMPLETE 50-TASK ROBOTWIN2.0 EVALUATION

  • AHS gains 6.80 percentage points over Base across all 50 RoboTwin2.0 tasks (task-cluster 95% interval [3.75, 9.95]); exact pairing covers only 62 of 100 task–setting cells.The full-matrix comparison retains all tasks, while the paired-episode interval applies only to the exact-identity subset.Source: §C.2 · p. 22

C.3 EIGHT-TASK POLICY-FAMILY EVALUATION

  • Fast-WAM success rises from 86.81% to 88.13% with AHS, while X-VLA rises from 47.44% to 47.75% overall, with Hard success decreasing from 20.38% to 19.75%.Individual task regressions are retained for both policies.Source: §C.3 · p. 22

C.1 QHA TRANSFER TO HELD-OUT TASKS

  • On two held-out tasks, QHA fusion improves success on two and reduces it on two; the aggregate gain over AHS does not establish uniform improvement or transfer of the base policy.The 400 AHS, QHA-only, and AHS+QHA episodes share identities, whereas Base results come from a separate evaluation.Source: §C.1 · p. 22, Table 7 · p. 23

C.4 ROBOCASA GR1 TABLETOP

  • On RoboCasa’s 24-task π0.5 evaluation, mean success is 40.08% for Base and 42.50% for AHS; the table also includes GR00T-family and other policy rows.QwenFAST and QwenPI have no AHS counterpart in this table.Source: §C.4 · p. 23

C.5 HORIZON-SELECTION COMPARISONS

  • The compact AHS-versus-Jerk-min comparison does not establish an SR advantage because its paired 95% difference interval includes zero.The study reports success and call-count trade-offs without exact compute matching.Source: §C.5 · p. 23
  • AAC-core obtains 52.58% success, but its Base/AHS references come from separate evaluations and do not establish paired differences or compute equivalence.AAC-core uses a joint-space adapter rather than an exact reproduction of the published Cartesian-action implementation.Source: §C.5 · p. 23

D.1 POLICY-CALL AND RUNTIME COSTS

The cost and latency analyses report implementation-specific runtime profiles and asynchronous execution comparisons. Reduced waiting with RTC-AHS does not imply uniformly higher success.

D.1 POLICY-CALL AND RUNTIME COSTS

  • Runtime means include failures, and the reported profiles are descriptive rather than hardware-matched cross-policy rankings.Episode wall time includes simulation and within-episode overhead, not physical robot execution time.Source: §D.1 · p. 23
  • Both QHA variants make fewer calls than the AHS reference but have higher policy-inference and wall times; the AHS total inference time is 3.49 s per episode.Success rates and these timing measurements come from different cohorts, so they cannot estimate success-normalized efficiency.Source: §D.1 · p. 25, Table 12 · p. 26
  • AAC-core averages 180.95 calls, 20.98 s total inference, and 47.56 s wall time per episode across 1,200 RoboCasa episodes.Each call samples 20 chunks; Base/AHS references are from separate evaluations and are not paired with these measurements.Source: §D.1 · p. 26

D.2 LATENCY AND ASYNCHRONOUS EXECUTION

  • At +200 ms, RTC reduces AHS waiting by 94.5% on Easy and 95.6% on Hard; RTC-AHS exceeds RTC-Fixed in aggregate SR in all six setting/delay conditions, but not uniformly versus Sync-AHS.RTC-AHS changes final SR versus Sync-AHS by −2.00 points on Easy and +3.00 on Hard at +200 ms; calls and model computation remain higher than RTC-Fixed.Source: Table 13 · p. 27, §D.2 · p. 27

E AHS HYPERPARAMETER SENSITIVITY

The sensitivity analyses vary AHS settings, while additional rollout examples illustrate phase-dependent horizons. The examples are qualitative rather than scored comparisons.

E AHS HYPERPARAMETER SENSITIVITY

  • Compared with Wh = 60, Wh = 30 and 90 yield gains of +2.73 and +3.91 percentage points, but both paired 95% intervals include zero; Wh = 60 is retained.The continuity-window sweep shares episode identities across 256 episodes per configuration.Source: E AHS HYPERPARAMETER SENSITIVITY · pp. 28–29, Figure 10 · p. 29
  • Across other AHS hyperparameter configurations, overall SR spans 34.38–37.13%, characterizing sensitivity rather than validation-based selection or an established optimum.One parameter family changes at a time, with the two penalties varied jointly.Source: E AHS HYPERPARAMETER SENSITIVITY · p. 29, Table 16 · p. 30

F MORE CASE STUDIES

  • In a successful π0 Handover Block rollout, AHS selects shorter horizons during grasping and transport, then longer horizons as motion stabilizes.This is an illustrative case, not a controlled comparison of posterior mechanisms.Source: Figure 11 · p. 30, F MORE CASE STUDIES · p. 30
  • Additional π0 RoboTwin2.0 and π0.5 real-world examples show horizon changes across phases, but are qualitative rather than additional scored trials.Real-world phase labels are manual visual annotations, not ground-truth boundaries or trial scores.Source: F MORE CASE STUDIES · p. 30, Figure 12 · p. 31, Figure 13 · p. 32

G DISCUSSION AND LIMITATIONS

ChunkTrust selects how much of a generated action chunk to execute, differing from methods that alter action generation or monitor execution after selection. Its evidence and evaluation are bounded by specific access and deployment constraints.

Discussion and limitations

  • AHS scores prefixes using one generated chunk’s trace and executed history, without conditional re-decoding or joint optimization of the generator’s inference schedule.Unlike observation-driven verification, it does not provide within-prefix monitoring for disturbances that arise after a chunk is selected.Source: G DISCUSSION AND LIMITATIONS · p. 33
  • ChunkTrust uses spectral variation and continuity with executed history to select execution length while leaving generated actions and base-policy weights unchanged.This differs from approaches that train or alter action representations, generation, or correction mechanisms.Source: G DISCUSSION AND LIMITATIONS · p. 33
  • AHS retains horizon preferences within an episode, while QHA learns a context-conditioned prior offline; neither retrieves rollout trajectories to steer action generation.ChunkTrust addresses horizon selection for existing chunk policies, not whether longer open-loop execution is intrinsically preferable to reactive control.Source: G DISCUSSION AND LIMITATIONS · p. 33
  • AHS requires action chunks and generation-time velocity traces, and replanning-time decisions cannot directly detect disturbances arising during the chosen prefix.Evaluation covers multiple policies, two simulation benchmarks, and four real-world bimanual tasks, not all embodiments, safety-critical tasks, or long-horizon mobile manipulation.Source: G DISCUSSION AND LIMITATIONS · p. 34
  • An incorrect horizon can commit a robot to unsafe motion outside the tested distribution, so deployment calls for limits, emergency stops, supervision, and task-specific validation.The paper names workspace and speed limits, emergency stops, and human supervision during evaluation.Source: G DISCUSSION AND LIMITATIONS · p. 34
  • The RoboTwin2.0 and RoboCasa benchmarks, policy checkpoints, and software remain subject to their original licenses, terms of use, and attribution requirements.Source: G DISCUSSION AND LIMITATIONS · p. 34
Loading 2609.39754v1…