PaperlayerView exact revision on arXiv

Source-linked AI summary

When to Switch: Reliable Action-Chunk Extension for Vision-Language-Action Models

Seonghoon Yu, Dongwon Kim, HyungRok Jung, Yoonjae Baek, Byung-kwan Lee, Suha Kwak, Jeany Son

arXiv:2610.05719v1cs.RO

TL;DR

Costly VLA inference creates pauses, while extending action chunks can make open-loop execution unreliable. RACE predicts subskill transition timing and conditions action generation on it. It improves success-rate comparisons at longer chunks and reduces real-robot idle time with 4× chunks.

  • Problem

    VLA inference is costly enough to create pauses, while extending action chunks trades fewer policy calls for less reliable open-loop prediction.

    Source: §1 · p. 1

  • Method

    RACE predicts per-action transition timing from an auxiliary one-step denoising pass and uses that prior to condition action generation.

    Source: §1 · p. 2, §3.2 · p. 4

  • Results

    At 2× longer chunks, RACE achieves higher success rates than recent state-of-the-art and efficient VLA approaches, and remains competitive at 4×.

    Source: §1 · p. 2

  • Takeaways & Limitations

    On a real robot, 4× longer chunks reduce idle time by about 5×, with higher success than fine-tuning at the same chunk length.

    Source: §1 · p. 2

    RACE success declines more steeply beyond Hexec=20 and falls below baseline at Hexec=40; near people, longer-chunk deployment requires additional safeguards.

    Source: ETHICS STATEMENT · p. 10, §F.1 · p. 31

Abstract

from arXiv · show

Vision-Language-Action (VLA) models serve as unified policies for robotic manipulation, yet their expensive inference forces robots to pause between policy calls, resulting in stop-and-go execution that interrupts smooth motion and prolongs task completion. Extending the action chunk reduces policy calls and hence these pauses, but predicting farther into the future makes long-chunk execution unreliable. To understand where this unreliability arises, we analyze action errors within long chunks and find that they concentrate around transitions between manipulation subskills, growing sharply with chunk length. This suggests the importance of transition timing, i.e., when to switch subskills within a chunk. Motivated by this observation, we introduce RACE (Reliable Action-Chunk Extension), a framework that predicts the transition timing from an auxiliary one-step denoising pass and conditions action generation on it. By learning and conditioning on transition timing, RACE reduces errors at subskill transitions and enables reliable execution of longer action chunks. Across simulation benchmarks, RACE outperforms fine-tuning at the same chunk length; with 2x longer chunks, it surpasses recent state-of-the-art and efficient VLAs in success rate, and with 4x longer chunks, it remains competitive. On a real robot, RACE uses 4x longer chunks, which reduces the idle time caused by stop-and-go execution by about 5x, while achieving a higher success rate than fine-tuning with the same chunk length. Code and a real-robot demo are available at https://github.com/Seonghoon-Yu/RACE-VLA

1 INTRODUCTION

Costly VLA inference creates pauses, while longer action chunks reduce policy calls but eventually become unreliable. RACE targets transition timing within chunks, where errors concentrate and grow, to support longer, more reliable execution.

1 INTRODUCTION

  • Extending action chunks reduces policy calls and pauses, but long open-loop predictions without intermediate feedback eventually make execution unreliable and lower success rates.Inference pauses accumulate into idle time that prolongs tasks and interrupts smooth motion.Source: §1 · p. 1
  • Action errors spike around subskill transitions, and those spikes grow substantially with chunk length as switches become harder to predict without feedback.The paper detects transitions as abrupt changes in action dynamics because demonstrations do not annotate subskills.Source: §1 · p. 2
  • RACE predicts transition timing from an auxiliary one-step denoising pass, then conditions every action-generation denoising step on that timing through token-wise modulation with learnable gates.The framework reduces transition error spikes across chunk lengths and improves task success at each chunk length.Source: §1 · p. 2
  • At 2× longer chunks, RACE achieves higher success rates than recent state-of-the-art and efficient VLA approaches, and remains competitive with them at 4×.On a real robot, 4× longer chunks reduce idle time by about 5×, with higher success than fine-tuning at the same chunk length.Source: §1 · p. 2

2 RELATED WORK

Related work spans VLA policy architectures, action chunking and asynchronous execution, efficient inference, and the use of subskill structure in policy learning. RACE differs by targeting transition timing within each chunk and making the full chunk reliable to execute.

2 RELATED WORK

  • VLA policies adapt pretrained vision-language models to robot control, with early models generating discrete action tokens and later models predicting continuous chunks or using flow-matching action experts.Examples include RT-2, OpenVLA, and later flow-matching approaches [Physical Intelligence et al., 2025; NVIDIA et al., 2025].Source: §2 · p. 3
  • Action chunking predicts multiple future actions per call, while adaptive methods execute reliable prefixes and PACE shortens execution around detected transitions; RACE aims to make the whole chunk reliable.PolicyTrim extends chunks through reinforcement learning, whereas RACE uses lightweight action-expert post-training; longer chunks can also complement asynchronous execution.Source: §2 · p. 3
  • Efficient VLA inference methods prune visual tokens or reuse computation across calls, leaving call counts unchanged, while RACE reduces calls; the approaches are described as orthogonal and combinable.The cited token-pruning methods include [Liu et al., 2025b], and computation reuse includes [Liu et al., 2026b].Source: §2 · p. 3
  • Prior subskill methods predict subskill goals, condition on inferred stages or progress, or jointly generate progress sequences; RACE predicts within-chunk transition timing and conditions generation on it.The cited approaches include change-point detection, manual stage annotation [Wu et al., 2025], and progress-sequence generation.Source: §2 · p. 3

3 RACE: RELIABLE ACTION-CHUNK EXTENSION

RACE adds a transition-timing prior to flow-matching action generation so VLAs can execute longer action chunks. It predicts the prior from an auxiliary denoising pass and uses it to condition action generation.

3.1 OVERVIEW OF RACE

  • RACE adds a per-action transition-timing prior to flow-matching chunk generation, then restarts full denoising from the same noise after an auxiliary pass.The VLM features are cached and reused for both passes.Source: §3.1 · pp. 3–4

3.2 TRANSITION-TIMING PREDICTION VIA ONE-STEP DENOISING

  • An auxiliary one-step pass yields action-position-aligned hidden states as coarse trajectory cues; the prediction head combines them with cached VLM features to score transition proximity across the chunk.The prior assigns a score to each action indicating how close it is to a subskill transition.Source: §3.2 · p. 4
  • Soft timing targets derive from PELT change points, and cross-entropy supervision also backpropagates into the action expert to encourage representations that capture upcoming transitions.The targets peak at detected transition points and decay with distance.Source: §3.2 · pp. 4–5

3.3 TRANSITION-CONDITIONED ACTION GENERATION

  • At each denoising step, learned gates scale transition embeddings into token-wise adaptive-RMSNorm offsets that modulate attention and feed-forward modules.The conditioning lets the action expert adapt denoising based on the predicted timing prior.Source: §3.3 · p. 5
  • Training combines full and auxiliary flow-matching losses with timing loss, conditions on jittered targets, and uses the predicted prior for conditioning at inference.Source: §3.3 · p. 5

4 EXPERIMENTS

RACE is evaluated across three manipulation benchmarks, through component and transition-timing analyses, and on a real robot. Across simulation benchmarks it outperforms fine-tuning at the same chunk length, while longer chunks reduce real-robot idle time.

4.1 EXPERIMENTAL SETUP

  • Evaluation covers VLABench, RoboCasa-H50, and LIBERO, measuring task success and inference-time speedup against π0.5 while matching execution chunk lengths for RACE and fine-tuned π0.5.The official π0.5 protocol predicts 10 actions and executes the first 5; Hexec denotes actions executed per policy call.Source: §4.1 · p. 5

4.2 MAIN RESULTS

  • Across VLABench and RoboCasa-H50, RACE outperforms fine-tuned π0.5 at every tested chunk length; at Hexec ≥10 it also beats efficient VLA methods in success rate and speedup.On LIBERO, RACE outperforms fine-tuned π0.5 at every Hexec and PolicyTrim at Hexec ∈{10, 15}.Source: §4.2 · pp. 6–7, Table 1 · p. 6, Table 2 · p. 7

4.3 ABLATION STUDY

  • Adding timing prediction and conditioning improves on fine-tuning, while the auxiliary denoising loss alone contributes little.Timing supervision raises success even without injecting the prediction into generation; conditioning adds a further gain, and Laux may provide better features for the prediction head.Source: §4.3 · p. 7, Table 3 · p. 7
  • PELT transition labels outperform speed-minimum labels, and combining VLM and action-expert features gives the prediction head its best reported result.Random transitions shrink modulation offsets by more than an order of magnitude; the combined input scores 33.0, versus 31.6 for VLM features and 32.5 for action features.Source: §4.3 · pp. 7–8, Table 4 · p. 8

4.4 UNDERSTANDING THE EFFECTIVENESS OF RACE

  • Transition-action errors exceed non-transition errors, and the gap widens at Hexec = 20; RACE reduces both, especially transition errors near the chunk end.RACE also places gripper switches within one step of the correct step more often, with a similar false-switch rate.Source: §4.4 · p. 8, Figure 4 · p. 8
  • RACE tends to gain more on tasks with more subskill transitions, and its positive advantage over fine-tuning grows with chunk length.The longer-chunk pattern is consistent with transition timing being an obstacle to chunk extension that RACE relieves.Source: §4.4 · pp. 8–9, Figure 5 · p. 8
  • Learned gates weight the prior most strongly at early denoising steps, while inference-time perturbations show that performance depends on the prior’s transition locations.Shifting the prior by one step lowers success below the zero-prior setting and shifts generated gripper switches in the same direction; the predicted prior does not capture every transition.Source: Figure 5 · p. 8, §4.4 · p. 9, Table 5 · p. 9

4.5 REAL-WORLD DEPLOYMENT

  • On a real robot, extending chunks reduces idle time because more actions per policy call allow the next chunk to be generated during execution; RACE also beats same-length fine-tuning in success rate.With asynchronous execution, task-completion time decreases only slightly because motion dominates it and most latency is already hidden.Source: §4.5 · p. 9

5 CONCLUSION

RACE uses transition timing to support longer action chunks, building on the finding that errors concentrate at subskill transitions and grow with chunk length. The appendix describes how those transitions are detected, counted, and represented as training targets.

5 CONCLUSION

  • RACE achieves competitive success rates while executing 4× longer chunks; on a real robot, it reduces idle time and beats same-length fine-tuning.The conclusion reports that longer chunks reduce real-robot idle time and yield higher success than fine-tuning at the same chunk length.Source: Table 6 · p. 9, §5 · p. 9

A.1 VISUALIZATION OF SUBSKILL TRANSITIONS

  • Detected change points align with visible subskill boundaries in VLABench examples; illustrated subskill names are not used as policy-learning labels.RACE uses transition points as labels, not the illustrated subskill categories.Source: Figure 7 · p. 15

A.3 CHANGE-POINT DETECTION

  • PELT segments each demonstration into subskills by detecting change points in standardized pose-difference and gripper-action signals.Per-dimension standardization prevents action dimensions with different scales from disproportionately influencing detection.Source: §A.3 · p. 16

A.4 SOFT LABELS FOR SUBSKILL TRANSITIONS

  • PELT outputs hard labels at individual steps, which are converted into per-position soft targets that peak at transitions and spread to neighboring actions.The prediction head estimates transition likelihood at each position in a chunk rather than a single transition index.Source: §A.4 · pp. 16–17

A.5 VALIDITY OF TRANSITION LABELS WITH DEMONSTRATION GENERATOR LABELS

The appendix evaluates transition labels against generator-recorded boundaries and adds comparisons on LIBERO and at longer execution lengths. RACE remains competitive as chunks extend, though success declines at the largest tested length.

A.5 VALIDITY OF TRANSITION LABELS WITH DEMONSTRATION GENERATOR LABELS

  • On 300 held-out VLABench demonstrations, PELT labels align more closely with generator-recorded subskill boundaries than random or speed-minimum labels.Precision and recall count transitions pooled across episodes, with a match defined by a ±3-step tolerance.Source: §A.5 · pp. 17–18, Table 8 · p. 18

B.1 COMPARISON ON LIBERO

  • At each tested Hexec, RACE beats same-length fine-tuned π0.5 on LIBERO; at Hexec ∈{10, 15}, it is competitive with recent VLAs and beats adaptive-chunking and efficient-VLA methods in success and speedup.At Hexec = 20, it remains competitive with the π0.5 baseline while requiring less inference time per episode.Source: §B.1 · p. 18, Table 9 · p. 19

B.2 SCALING TO LONGER CHUNKS

  • RACE beats the π0.5 baseline through Hexec = 30 but falls slightly below it at Hexec = 40.Longer chunks reduce per-episode inference time, while success declines at larger execution lengths; the authors attribute the decline likely to open-loop drift not fully compensated by the timing prior.Source: Table 10 · p. 19, §B.2 · p. 19

B.3 RESULTS WITH VLM JOINT TRAINING

The appendix tests joint VLM training, prediction length, other VLA baselines, latency, and training cost. RACE’s advantages persist across several settings, while its added inference and training costs remain modest relative to the compared methods.

B.3 RESULTS WITH VLM JOINT TRAINING

  • Joint VLM training improves average success rate and progress score, with largest gains on instruction and texture tracks, but requires substantially more memory and compute.The main experiments therefore keep the VLM frozen.Source: §B.3 · p. 19, Table 11 · p. 20, Table 12 · p. 20

B.4 DECOUPLING PREDICTION AND EXECUTION CHUNK LENGTHS

  • Predicting 40 actions while executing the first 20 improves both RACE and fine-tuned π0.5 on VLABench.The authors suggest that executing only the first half avoids the less accurate actions near the chunk’s end.Source: Table 13 · p. 20, §B.4 · p. 20

B.5 GENERALIZATION TO OTHER VLA BASELINES

  • On LIBERO, RACE beats fine-tuned variants at every chunk length on GR00T-N1.6 and OpenVLA-OFT, but only GR00T-N1.6 recovers or surpasses the baseline.On OpenVLA-OFT, the gain is smaller; the authors suggest the absence of iterative denoising may be one reason.Source: §B.5 · p. 20, Table 14 · p. 21

B.6 LATENCY OVERHEAD COMPARISON

  • RACE’s added modules cost about 4% of a policy call, while ACoT-VLA’s explicit action-reasoner denoising alone costs about as much as the action expert.The authors attribute ACoT-VLA’s lower speedup than RACE at the same Hexec to this latency gap.Source: Table 15 · p. 21, §B.6 · p. 21

B.7 TRAINING EFFICIENCY COMPARISON

  • RACE’s training cost is moderately higher than fine-tuned π0.5 and far lower than ACoT-VLA or PolicyTrim.Across four LIBERO suites, PolicyTrim is estimated to take 1,828 hours, over 100 times longer than RACE.Source: §B.7 · p. 21, Table 17 · p. 22, Table 16 · p. 22

B.8 STATISTICAL SIGNIFICANCE

  • Across three RACE runs on VLABench at Hexec = 20, every run outperforms the once-trained fine-tuned π0.5 baseline in Tab. 1 by a margin much larger than run-to-run variation.Individual tracks vary more across runs, so small single-track differences warrant care. The Tab. 1 model is Run 1, whose results are close to the mean.Source: Table 18 · p. 22, §B.8 · p. 22

C ADDITIONAL ABLATION STUDY

Ablations find stable performance across PELT settings and modest benefits from learnable gates and soft labels. Further analyses show that the timing head misses some transitions, while shifts to its prior affect success and generated switching times.

C.1 SENSITIVITY TO CHANGE-POINT DETECTION SETTINGS

  • Performance stays stable as the PELT penalty varies from β = 2 to 8, despite a threefold change in label density.The authors conclude that RACE is not sensitive to these PELT settings.Source: Table 19 · p. 22, §C.1 · p. 22

C.2 EFFECT OF PER-STEP LEARNABLE GATE

  • Replacing the per-step learnable gates with a fixed αk = 1 performs slightly worse than allowing the conditioning strength to vary by denoising step.The learned gates take on a step-dependent pattern after training.Source: Table 20 · p. 23, §C.2 · p. 23

C.3 EFFECT OF SOFT TRANSITION LABEL

  • Hard one-hot transition labels lower success relative to soft labels spread across neighboring steps.The authors note that transitions span several actions and detected points may not exactly match the true transition steps.Source: Table 21 · p. 23, §C.3 · p. 23

C.4 EFFECT OF TEACHER FORCING AND JITTERING

  • Teacher forcing and jittering each have only a marginal effect, while jittering mitigates performance drops from inference-time shifts of several steps.During training, jittering randomly shifts the conditioning target by one action step with probability 0.5.Source: Table 22 · p. 23, §C.4 · p. 23

D.1 ACCURACY OF TRANSITION-TIMING PREDICTION

  • The timing head predicts transitions with high precision but misses some; its evaluation covers demonstration states, not policy-rollout states.On held-out demonstrations, accuracy drops most for localization with longer chunks; multi-transition recall is lower than precision.Source: §D.1 · pp. 23–24, Table 23 · p. 24

D.2 SENSITIVITY TO TRANSITION-TIMING SHIFT AT INFERENCE

  • Any inference-time shift lowers success, even by one step; jittering mitigates the drop for shifts of several steps.At the largest shifts, partial recovery reflects peaks dropped beyond the chunk, not tolerance to the shift.Source: §D.2 · pp. 24–25, Table 24 · p. 25

D.3 EFFECT OF THE PRIOR ON GENERATED SWITCHING TIMING

  • Shifting the prior moves generated gripper switches in the same direction, but by only about a quarter as many steps.Replacing the prior with zeros leaves switch offset nearly unchanged but increases missed switches.Source: Table 25 · p. 25, §D.3 · p. 25

D.4 DECOMPOSING TRANSITION ERRORS BY ACTION COMPONENT

Action errors are decomposed by component, and gripper-switch timing is evaluated directly. RACE reduces transition errors across components and improves several gripper-switch timing measures over fine-tuned π0.5.

D.4 DECOMPOSING TRANSITION ERRORS BY ACTION COMPONENT

  • Component errors are measured at positions within ±1 step of PELT change points, using 500 training demonstrations and four noise seeds.Errors are split into translation (x, y, z), rotation (roll, pitch, yaw), and gripper components.Source: §D.4 · p. 25
  • RACE reduces transition errors in translation, rotation, and gripper components relative to fine-tuned π0.5.Translation has the largest errors and dominates overall error; reductions are larger in rotation and gripper. The analysis uses 500 VLABench training demonstrations at Hexec = 20 and averages over four noise seeds.Source: §D.4 · p. 25, Table 26 · p. 26
  • RACE more often switches within ±1 and ±3 steps and misses fewer switches than fine-tuned π0.5, with a similar false-switch rate.Exact-step switching improves on training demonstrations, but not significantly on held-out demonstrations; mean error is lower on held-out data but not significantly so.Source: §D.4 · p. 26, Table 27 · p. 26

D.5 DECOMPOSING IDLE TIME IN REAL-ROBOT EXPERIMENTS

In asynchronous execution, the robot continues moving while a new chunk is prepared and pauses only when inference outlasts the remaining-action buffer. Longer chunks can hide latency, although fixed startup and control-loop costs remain.

D.5 DECOMPOSING IDLE TIME IN REAL-ROBOT EXPERIMENTS

  • The execution buffer is the time needed to execute the actions remaining in a chunk when it arrives, after real-time chunking skips actions elapsed during inference.The robot pauses only if policy-call latency exceeds this buffer.Source: §D.5 · p. 26
  • With Hexec = 5, about three of ten actions are skipped on arrival, leaving a roughly seven-action buffer that is slightly shorter than average call latency; pauses account for most idle time.The robot therefore executes about seven actions per call rather than five. The table decomposes idle time over successful trials and defines pauses by latency exceeding the execution buffer.Source: §D.5 · p. 26, Table 28 · p. 27
  • With Hexec = 20, the buffer is several times longer than latency, so no pause occurs after the first call; startup and control-loop overhead remain.The chunk length needed to remove pauses depends on deployment setup, including per-call latency such as camera capture; RACE's additional latency here reflects an implementation without CUDA graphs.Source: §D.5 · p. 27

E.1 DETAILS ON REAL-ROBOT EXPERIMENTS

The appendix details the real-robot evaluation, simulation speedup measurement, and adaptation of RACE to other VLA backbones. It reports that simulation speedup reflects fewer policy calls rather than shorter episodes, while noting that speedup should be read alongside success rate.

E.1 DETAILS ON REAL-ROBOT EXPERIMENTS

  • The real-robot evaluation uses a bowl pick-and-place task, 30 training demonstrations, and 50 trials per method, reporting success, completion time, and idle time.All methods use asynchronous execution with real-time chunking and predict chunks twice as long as they execute.Source: §E.1 · pp. 27–28
  • Real-robot policies are fine-tuned from π0.5 on 30 demonstrations; RACE uses H = 40, compared with fine-tuned π0.5 at H = 10 and H = 40.Policies are evaluated on an AgileX PiPER arm with wrist and front cameras; deployment requests a new chunk after H/2 actions.Source: §E.1 · pp. 27–28

E.2 DETAILS ON SIMULATION SPEEDUP MEASUREMENT

  • Simulation speedup compares baseline and model inference time per episode using single-call latency and average calls over commonly successful episodes.The measure excludes simulator execution time, accounts for per-call latency as well as call count, and is largely determined by Hexec and latency.Source: §E.2 · p. 28
  • At Hexec = 20, the three models take similar environment steps per episode, so speedup comes from fewer policy calls rather than shorter episodes.RACE makes slightly more calls than fine-tuned π0.5 and adds a small per-call overhead, so its speedup is slightly lower; the breakdown covers 127 episodes completed by all three models.Source: §E.2 · p. 28, Table 29 · p. 29
  • Speedup remains similar across different episode sets, including all episodes regardless of success, and should be read with success rate, where RACE outperforms fine-tuned π0.5.Table 30 reports speedup under different episode selections; its first row is the main-comparison setting.Source: Table 30 · p. 29, §E.2 · p. 29

E.3 DETAILS ON OTHER VLA BASELINES

  • RACE transfers to GR00T-N1.6 by injecting its prior through AdaLN layers, and to OpenVLA-OFT through RMSNorm layers in a modulated action-token pass.OpenVLA-OFT has no denoising loop, so its adaptation uses a single modulated pass without a per-step gate.Source: §E.3 · p. 29

E.4 DETAILS ON ACTION ERROR MEASUREMENT

The analysis measures demonstration-state action errors around PELT-detected transitions and compares them with matched non-transition positions. The transition error gap persists after matching on chunk position and a measure of local action variation.

E.4 DETAILS ON ACTION ERROR MEASUREMENT

  • Action error is the normalized-action-space MSE against demonstration actions, averaged across seven dimensions; the comparison locates errors but does not isolate causes of rollout failures.Ground-truth actions are available only at demonstration states, not at states visited during rollouts.Source: §E.4 · p. 29
  • For transition-aligned errors, actions are averaged at offsets −8 to 8 from each PELT change point; within-chunk analysis separates transition from non-transition actions.Each change point contributes separately, and only offsets within the chunk are retained.Source: §E.4 · pp. 29–30
  • Transition errors remain about 1.3 times matched non-transition errors for both π0.5 and RACE.Matching uses chunk position and deciles of second-order action difference; non-transition positions are at least three steps from any change point.Source: §E.4 · p. 30, Table 31 · p. 30

E.5 DETAILS ON DATASET AND EVALUATION PROTOCOLS

E.5 defines the datasets and evaluation protocols used for VLABench, RoboCasa-H50, and LIBERO. E.6 specifies RACE’s two-pass training and inference procedures.

E.5 DETAILS ON DATASET AND EVALUATION PROTOCOLS

  • VLABench evaluation averages success rate and progress score across ten tasks and five tracks, using 50 episodes per task per track.Its 5,000 training demonstrations comprise 500 scripted demonstrations per task; the five tracks vary distribution, categories, commonsense demands, instructions, and textures.Source: §E.5 · p. 30
  • RoboCasa-H50 evaluates 25 kitchen-manipulation tasks with 50 episodes each and weights category success rates by the number of tasks per category.The training set contains 50 human-teleoperated demonstrations per task, totaling 1,250.Source: §E.5 · p. 30
  • LIBERO evaluation uses 50 episodes per task across four ten-task suites, reporting each suite’s mean success rate and the average across suites.Training uses 50 human-teleoperated demonstrations for each of 40 tasks, totaling 2,000.Source: §E.5 · p. 30

E.6 PSEUDOCODE FOR TRAINING AND INFERENCE

  • RACE training first predicts transition timing in an unmodulated auxiliary pass, then trains the full pass with a jittered timing target and a denoising-step gate.The full-pass action expert is trained at a sampled flow time; inference predicts the prior once per policy call and applies the corresponding gate at each denoising step.Source: §E.6 · pp. 30–31

F.1 LIMITATIONS

RACE’s transition-timing predictions can miss transitions, and its success declines beyond longer chunk lengths. The broader-impact discussion also limits the evidence to simulation and controlled laboratory settings and calls for safeguards in real-world deployment.

F.1 LIMITATIONS

  • RACE’s timing head can miss transitions, particularly on held-out demonstrations and longer chunks, and its timing labels come from change-point detection rather than human annotations.Its success rate declines more steeply beyond Hexec = 20 and falls below the baseline at Hexec = 40; the authors attribute this to open-loop error growth that transition conditioning does not reduce.Source: §F.1 · p. 31
  • RACE is designed for iterative denoising, and its gain is smaller on a VLA without a denoising process.The authors nevertheless report that it outperforms fine-tuned variants at every chunk length on GR00T-N1.6 and OpenVLA-OFT.Source: §F.1 · p. 31

F.2 BROADER IMPACT

  • Experiments are limited to simulation benchmarks and a controlled laboratory setting; real-world deployment, especially near people, requires additional safeguards.The authors name safety monitoring, collision checking, and human supervision.Source: §F.2 · p. 32
  • Longer chunks leave fewer opportunities to react to unexpected environmental changes, while demonstration-based training may inherit unsafe behaviors or biases.The discussion gives a person entering the workspace as an example of an unexpected change.Source: §F.2 · p. 32

Related papers

Loading 2610.05719v1…