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

cs.ROSep 30, 2026

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

Authors: Fanding Huang, Jingyan Jiang, Shifeng Bao, Mingkang Pu, Shiwei Li, Jing Xu, Shijia Xu, Guanbo Huang, +8 more

Organizations: Tsinghua University · Beijing Academy of Artificial Intelligence (BAAI) · Shenzhen Technology University · Renmin University of China · Hefei University of Technology · Jiangnan University · Chongqing University · The Chinese University of Hong Kong

Abstract

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 $π_{0.5}$ 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 $π_{0.5}$ 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/

Figures & tables

Figure 1: Fixed execution horizons are brittle across tasks and settings. Top: the reliable executable prefix is phase dependent in a real bread-manipulation rollout (qualitative trust illustration $K_{\mathrm{exec}}=K_{t}$ ). (a) A shared robot policy with fixed or adaptive execution horizons. (b) Fixed horizons $K\in\{10,20,30,40,50\}$ versus adaptive execution on $\pi_{0}$ RoboTwin2.0 evaluations, grouped by task length and environment setting. Error bars indicate $\pm 1$ standard error of the mean across task–setting combinations. Dashed lines show the corresponding adaptive references.

Figure 2: Horizon-aware evidence and episode outcomes. Analysis of 1,600 RoboTwin2.0 episodes with

\pi_{0}

and fixed

K=H=50

. (a,b) Episode means of replan-level evidence:

\bar{z}_{\mathrm{intra}}(k)=R_{z}^{-1}\sum_{t}z_{\mathrm{intra},t}(k)

\bar{u}_{\mathrm{inter}}(k)=R_{u}^{-1}\sum_{t}u_{\mathrm{inter},t}(k)

. Sums and counts

R_{z},R_{u}

use valid replans for each metric and horizon. Lines/bands show medians/interquartile ranges across episodes. (c) Failure rates after median-splitting episode risks (within-episode 75th percentiles of

1-q_{\mathrm{intra}}

and

1-q_{\mathrm{inter}}

). “Inter only”/“Intra only” means only the named risk is high. Details: Appendix B .

Figure 3: Overview of horizon adaptation. AHS normalizes raw evidence, maintains per-candidate Beta reliability, and forms selection distribution $\mu_{t}$ . QHA learns a context-conditioned prior from dense evidence.

Policy	Task Scope	Train Recipe	Base	+AHS	$\Delta$ (pp)
RoboTwin2.0 Easy and Hard
$\pi_{0.5}$	50 tasks	multitask post-training	56.70	63.50	+6.80
$\pi_{0}$	8 tasks	task-specific post-training	18.19	21.38	+3.19
$\pi_{0.5}$	8 tasks	task-specific post-training	29.63	36.25	+6.63
Fast-WAM	8 tasks	multitask post-training	86.81	88.13	+1.31
RoboCasa GR1 Tabletop

Table 1: AHS performance across simulation benchmarks. Success rate (%). Each comparison uses the same policy checkpoint.

\Delta=\mathrm{SR}_{\mathrm{AHS}}-\mathrm{SR}_{\mathrm{Base}}

A. QHA training-task evaluation
	$\pi_{0}$ ( Black et al., 2024 )						$\pi_{0.5}$ ( Intelligence et al., 2025 )
Task	Base		AHS		AHS+QHA		Base		AHS		AHS+QHA
	Easy	Hard	Easy	Hard	Easy	Hard	Easy	Hard	Easy	Hard	Easy	Hard
Blocks Ranking RGB	19.00	0.00	28.00	1.00	31.00	5.00	36.00	20.00	48.00	29.00	56.00	33.00
Handover Block	41.00	10.00	41.00	5.00	64.00	9.00	44.00	14.00	44.00	14.00	32.00	10.00
Handover Mic	100.00	2.00	100.00	24.00	100.00	30.00	98.00	64.00	100.00	56.00	98.00	59.00

Table 2: QHA augmentation and transfer on RoboTwin2.0. Success rate (%). Easy and Hard denote clean and randomized settings. A: Evaluation on the eight tasks used to train QHA. B: For

\pi_{0.5}

, QHA is trained on six tasks and evaluated on two held-out tasks. Results are averaged over Easy and Hard. Bold marks the best method for each policy and evaluation condition.

Panel of Figure 4 — Figure 4: Real-world deployment. (a) Four bimanual household tasks: fold towels, move bread to a plate, move a drink to a basket, and put a duck toy into a drawer. (b) Mean normalized process scores for $\pi_{0.5}$ and $\pi_{0.5}$ + AHS over 15 rollouts per task and method. Error bars show $\pm$ one standard error of the mean across rollouts.

Candidate set	$\|\mathcal{K}\|$	Success (%)	$\Delta$ Success	Replan time (ms)	Overhead (%)	Replan counts	Ep. time (s)
Fixed	–	22.00	–	95.83	–	4.22	22.43
$\mathcal{K}_{5}$	5	33.00	+11.00	96.50	0.70	8.36	23.17
$\mathcal{K}_{10}$	10	32.50	+10.50	96.84	1.05	9.34	23.08
$\mathcal{K}_{50}$	50	34.88	+12.88	99.28	3.60	10.57	21.39

Table 3: Candidate-set scaling on

\pi_{0.5}

. Success and per-replan overhead over four RoboTwin2.0 tasks as

|\mathcal{K}|

increases. Fixed denotes

\pi_{0.5}

without horizon selection.

\Delta

Success is in percentage points.

Method	Easy (%)	Hard (%)	Overall (%)
Base $\pi_{0}$ (default)	20.75	4.00	12.38
Inter-chunk only	23.50	4.00	13.75
Intra-chunk only	23.75	5.25	14.50
Both evidence terms, no posterior	25.50	3.75	14.63
Full AHS	24.25	5.75	15.00
Full AHS+QHA	27.50	4.00	15.75

Table 4: Component ablation (

\pi_{0}

). Four-task mean success. Top two: bold/underline.

Figure 5: AHS with asynchronous execution on

\pi_{0.5}

Hard. (a) Mean waiting and observation age. Points run left to right as +0/+100/+200 ms additional delay. (b) Final SR. (c) SR(

t

) at +200 ms. All 400 outcomes per condition are included. Bars and shading are marginal/pointwise 95% paired-seed bootstrap intervals within four tasks. Time uses a controlled physical clock with a 142 ms base delay, not native deployment timing. Easy curves and full results: Appendix D.2 .

Appendix figures & tables20 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: Real-world robot platform. AgileX COBOT Magic configured as an ALOHA-style bimanual system for teleoperation, data collection, and autonomous policy execution.

Task and sub-step	0	0.5	1
Bread: grasp	Fails to grasp	Multiple attempts	Smooth first attempt
Bread: handover	Handover fails	Unstable/awkward receiving grasp	Stable, aligned grasp
Bread: place on plate	Not placed	Poor alignment or rough placement	Clean placement
Drink: push	Only tilts, no useful displacement	Acceptable position, tilt/misalignment	Good position, stable alignment
Drink: grasp	Fails to grasp	Multiple attempts	Smooth first attempt
Drink: place in basket	Fails or drops drink	Rough placement/collision	Clean placement

Appendix

Table 5: Real-world scoring criteria. Scoring is independent of arm assignment.

Figure 7: Full evidence distributions and boundary behavior. (a,b) Episode means of the operational scores, with violin marks indicating medians. Widths show within-group density, not relative sample counts. (c,d) Normalized speed profiles for candidate

k=10

and

k=50

, with dashed vertical lines marking the history–prediction join. Curves show means and bands show P10–P90 across episode profiles, not confidence intervals or the IQR bands used in the main figure.

Figure 8: Joint-risk landscape on the same 1,600 episodes. Dashed lines mark the original median thresholds. Points denote episodes, with outcome colors as in Fig. 7 . Color shows the smoothed local failure fraction. Lower opacity denotes lower smoothed occupancy. The color scale saturates below 0.25 and above 0.95. The map is descriptive. Group failure rates below are computed directly from episode outcomes, not read from the smoothed colors.

Group	Intra risk	Inter risk	Episodes	Failures	Failure rate
Both low	Low	Low	313	157	50.2%
Inter only	Low	High	193	144	74.6%
Intra only	High	Low	487	417	85.6%
Both high	High	High	607	591	97.4%

Appendix

Table 6: Exact median-split group outcomes. Counts and failure rates reproduce Fig. 2 (c).

Task	Setting	AHS	QHA-only	AHS+QHA	$\Delta$ (pp)
Blocks Ranking RGB	Easy	44.00	48.00	57.00	+13.00
Blocks Ranking RGB	Hard	33.00	29.00	30.00	-3.00
Place Bread Basket	Easy	52.00	50.00	49.00	-3.00
Place Bread Basket	Hard	31.00	31.00	35.00	+4.00
Overall	Both	40.00	39.50	42.75	+2.75

Appendix

Table 7: Complete held-out comparison. Success rate (%) for the

\pi_{0.5}

policy with QHA trained on six tasks and evaluated on two held-out tasks. Each task and setting uses 100 paired episodes.

\Delta=\mathrm{SR}_{\mathrm{AHS+QHA}}-\mathrm{SR}_{\mathrm{AHS}}

Task	Easy		Hard		$\Delta$ (pp)
	Base	AHS	Base	AHS
adjust bottle	100.00	100.00	90.00	95.00	+2.50
beat block hammer	70.00	80.00	15.00	50.00	+22.50
blocks ranking rgb	80.00	95.00	40.00	70.00	+22.50
blocks ranking size	25.00	45.00	25.00	35.00	+15.00
click alarmclock	60.00	65.00	45.00	45.00	+2.50

Appendix

Table 8: Complete 50-task RoboTwin2.0 results. Success rate (%) for the shared multitask

\pi_{0.5}

checkpoint, with 20 episodes per task, setting, and method.

\Delta

is AHS minus Base, averaged over Easy and Hard.

	Fast-WAM ( Yuan et al., 2026 )				X-VLA ( Zheng et al., 2026 )
Task	Base		AHS		Base		AHS
	Easy	Hard	Easy	Hard	Easy	Hard	Easy	Hard
Blocks Ranking RGB	99.00	98.00	100.00	100.00	87.00	34.00	95.00	38.00
Handover Block	94.00	81.00	91.00	82.00	89.00	2.00	88.00	1.00
Handover Mic	100.00	99.00	100.00	100.00	97.00	1.00	100.00	1.00
Hanging Mug	66.00	64.00	71.00	69.00	32.00	8.00	35.00	4.00

Appendix

Table 9: Fast-WAM and X-VLA per-task success rates on RoboTwin2.0. Success rate (%) over 100 rollouts per task and setting. Easy and Hard denote clean and randomized settings. Each Base–AHS comparison fixes the policy checkpoint. Yellow columns use AHS, and bold marks the better method within each policy and setting, including ties.

Task	QwenFAST +Qwen3VL	QwenPI +Qwen3VL	Isaac-GR00T N1.5		Isaac-GR00T N1.6		Qwen3GR00T +Qwen3VL		$\boldsymbol{\pi_{0.5}}$
			Base	AHS	Base	AHS	Base	AHS	Base	AHS
PnPBottleToCabinetClose	38.0	26.0	64.0	68.0 (+4.0)	51.5	54.0 (+2.5)	46.0	64.0 (+18.0)	64.00	68.00 (+4.00)
PnPCanToDrawerClose	44.0	62.0	18.0	12.0 (-6.0)	13.0	12.0 (-1.0)	80.0	80.0 (+0.0)	58.00	56.00 (-2.00)
PnPCupToDrawerClose	56.0	42.0	12.0	4.0 (-8.0)	8.5	14.0 (+5.5)	54.0	52.0 (-2.0)	34.00	40.00 (+6.00)
PnPMilkToMicrowaveClose	44.0	50.0	38.0	34.0 (-4.0)	14.0	20.0 (+6.0)	48.0	42.0 (-6.0)	40.00	44.00 (+4.00)
PnPPotatoToMicrowaveClose	14.0	42.0	54.0	36.0 (-18.0)	41.5	50.0 (+8.5)	28.0	28.0 (+0.0)	22.00	30.00 (+8.00)

Appendix

Table 10: Full per-task success rates on RoboCasa GR1 Tabletop. Full 24-task breakdown, with 50 rollouts per task for

\pi_{0.5}

. Yellow columns use AHS and parentheses report percentage-point deltas over Base. Bold marks the best result in each row, including ties.

RoboTwin2.0
Selector	SR (%) $\uparrow$	Calls/ep	Policy infer	Total infer	Wall
Global fixed ( $K=20$ )	14.84	25.96	2.75	2.75	54.33
Jerk-min	14.06	21.14	2.28	2.28	52.01
EMA $q_{\mathrm{mix}}$	10.94	19.72	2.05	2.07	46.72
AHS (Full Beta)	15.63	20.71	2.26	2.28	53.47
Instantaneous $q_{\mathrm{mix}}$ (reference)	13.28	19.58	—	—	—

Appendix

Table 11: Horizon-selection success rates and computational costs. RoboTwin2.0 uses

\pi_{0}

on four tasks under Easy and Hard (128 episodes per selector). Instantaneous is a separate reference without timing measurements. Inference and wall times are seconds per episode. RoboCasa references are non-paired. Bold/underline mark best/second-best SR among the displayed methods within each benchmark.

	Calls/	Policy inference	Episode wall
Method	episode	(s/episode)	time (s/episode)
A. Full 50-task evaluation 2,000 episodes per method
Base	8.00	0.97	38.19
AHS	14.50	1.55	36.83
B. QHA timing evaluation 2 held-out tasks, 400 episodes per method
AHS	33.77	3.29	68.60

Appendix

Table 12: Policy-inference and episode costs for

\pi_{0.5}

. RoboTwin2.0 means include failures. Panels use separate cohorts. Panel B also differs from the success-rate evaluation in Tab. 2 B.

All four tasks
Method	SR (%) $\uparrow$			Wait (%) $\downarrow$	Wait s/ep $\downarrow$	Calls/ep	Infer s/ep	Mean age (s) $\downarrow$
	+0 ms	+100 ms	+200 ms
Sync-Fixed	19.50	19.50	19.25	3.77	4.432	12.89	5.04	5.02
Sync-AHS	26.25	24.25	23.75	6.80	7.893	22.95	8.59	3.00
RTC-Fixed	23.00	21.50	22.00	0.31	0.344	13.32	6.47	4.92
RTC-AHS	26.50	25.75	26.75	0.31	0.344	24.30	10.62	3.07

Appendix

Table 13: Latency results by task and in aggregate on RoboTwin2.0 Hard. 100 paired episodes per task/method/delay, including failures. Costs and mean observation age use +200 ms. Fixed uses target

K=40

. Bold/underline mark best/second-best SR, waiting, and age within each group, including displayed ties. Calls and inference time are unranked.

All four tasks
Method	SR (%) $\uparrow$			Wait (%) $\downarrow$	Wait s/ep $\downarrow$	Calls/ep	Infer s/ep	Mean age (s) $\downarrow$
	+0 ms	+100 ms	+200 ms
Sync-Fixed	38.00	35.50	37.00	3.90	3.856	11.21	4.38	5.12
Sync-AHS	46.00	46.25	44.75	6.63	6.310	18.35	6.90	3.16
RTC-Fixed	38.75	40.25	39.00	0.37	0.344	11.50	5.57	5.00
RTC-AHS	46.00	48.50	42.75	0.37	0.344	20.32	8.80	3.18

Appendix

Table 14: Latency results by task and in aggregate on RoboTwin2.0 Easy. 100 paired episodes per task/method/delay, including failures. Costs and mean observation age use +200 ms. Fixed uses target

K=40

. Bold/underline mark best/second-best SR, waiting, and age within each group, including displayed ties. Calls and inference time are unranked.

Figure 9: AHS with asynchronous execution on

\pi_{0.5}

Easy. The three panels mirror Fig. 5 : waiting versus mean observation age at +0/+100/+200 ms, final SR at each delay, and SR(

t

) at +200 ms. All 400 outcomes per condition are included. Bars and shading show marginal/pointwise 95% paired-seed bootstrap intervals within four tasks.

Easy
Delay (ms)	Method	Wait (%) $\downarrow$	Wait s/ep $\downarrow$	Calls/ep	Infer s/ep	Mean age (s) $\downarrow$
0	Sync-Fixed	1.67	1.600	11.11	4.32	4.93
	Sync-AHS	2.87	2.614	18.15	6.78	3.01
	RTC-Fixed	0.16	0.147	11.20	5.38	5.00
	RTC-AHS	0.16	0.147	17.86	7.81	3.24
100	Sync-Fixed	2.82	2.755	11.29	4.37	4.98

Appendix

Table 15: Aggregate costs at +0 and +100 ms on Easy and Hard. Each method/delay has 400 outcomes per setting, including failures. The +200 ms aggregates appear with the per-task results in Tabs. 13 and 14 . Observation age averages within episodes, then equally across episodes and tasks. Within each setting and delay, best and second-best distinct waiting and age values are bold and underlined. Displayed ties share a mark. Calls/ep and Infer s/ep are descriptive quantities and are not ranked.

Figure 10: Inter-chunk continuity-window sensitivity.

\pi_{0.5}

+ AHS on eight RoboTwin2.0 tasks, Easy and Hard, with 256 matched episodes per configuration. Lines connect tested points. Paired difference intervals appear in the text.

Parameter	Default	Tested	Easy $\uparrow$	Hard $\uparrow$	Overall $\uparrow$
Forgetting $\rho_{f}$	0.99	0.95	46.88	26.75	36.81
		0.995	46.50	25.88	36.19
Update strength $\eta$	1	0.5	45.38	24.38	34.88
		2	47.38	26.88	37.13
Kernel bandwidth $\sigma_{K}$	10	5	45.50	26.13	35.81
		20	44.25	24.50	34.38

Appendix

Table 16: Sensitivity to other AHS hyperparameters. Success rates (%) on eight RoboTwin2.0 tasks, with 1,600 episodes per configuration. Bold/underline mark higher/lower SR within each parameter family. Defaults specify parameters, not additional evaluations.

Figure 11: $\pi_{0}$ Handover Block case study. The selected execution horizon changes with task phase in a successful RoboTwin2.0 rollout. $K_{\mathrm{exec}}$ denotes the selected $K_{t}$ .

Figure 12: Additional $\pi_{0}$ case studies on RoboTwin2.0. Four successful rollouts show phase-dependent AHS horizons across Place Bread Basket and Blocks Ranking RGB under Easy and Hard settings.

Figure 13: Real-world $\pi_{0.5}$ +AHS case studies. Highlighted observations are selected at large adjacent-replan changes in $K_{t}$ . Curves show all recorded execution horizons without smoothing, through the end of each recording. Phase labels are manual visual annotations, not ground-truth boundaries or trial scores.

Explore similar work

Dynamic Execution Horizon Prediction for Chunk-based Robot Policies
Jun 9, 2026Yuchi Zhao, Miroslav Bogdanovic, Arjun Sohal +5Action Chunks Long-Horizon Manipulation
Department of Computer Science, University of Toronto, 40 St George St., Toronto, ON M5S 2E4, Canada · Vector Institute for Artificial Intelligence, W1140-108 College St., Schwartz Reisman Innovation Campus, Toronto, ON M5G 0C6, Canada · Acceleration Consortium, 700 University Ave., Toronto, ON M7A 2S4, Canada +7
PACE: Phase-Aware Chunk Execution for Robot Policies with Action Chunking
May 30, 2026Junnan Nie, Jiayi Li, Chenghao Liu +5Action Chunks Robot Policies
Peking University · Peking University. · JD Explore Academy. +1
Continue or Replan? Bernoulli-Continuation Policy Learning for Adaptive Horizon Execution
Aug 4, 2026Weichen Xu, Zhenhua Liu, Lin Luo +8Diffusion-Based Vision-Language-Actions Iterative Re-Planning
1Peking University · 2Microsoft Research Asia

Jun 9, 2026, cs.ROJ/K move · Enter open · S save

Dynamic Execution Horizon Prediction for Chunk-based Robot Policies

Yuchi Zhao, Miroslav Bogdanovic, Arjun Sohal, Liyu Tao+4

Department of Computer Science, University of Toronto, 40 St George St., Toronto, ON M5S 2E4, Canada · Vector Institute for Artificial Intelligence, W1140-108 College St., Schwartz Reisman Innovation Campus, Toronto, ON M5G 0C6, Canada · Acceleration Consortium, 700 University Ave., Toronto, ON M7A 2S4, Canada+7

Action Chunks Long-Horizon Manipulation Robot Policies Action Prediction Action Generation Short-Horizon Forecasting+1

Open paper AskPDF arXiv

Action chunking has become a standard design in modern robot policies, from diffusion/flow policies to vision-language-action models, where the policy predicts a sequence of actions and executes a fixed number of them instead of acting one step at a time. However, this paradigm relies on a key assumption: a fixed execution horizon. During chunk execution, the policy operates open-loop, which is particularly problematic for fine-grained manipulation tasks that require frequent replanning. In practice, the execution horizon is typically chosen through empirical tuning and is highly task-dependent. To this end, we propose Dynamic Execution Horizon Prediction (DEHP), an effective method that trains a lightweight execution-horizon prediction branch using online reinforcement learning while keeping the pretrained chunk policy completely frozen. This makes the method compatible with black-box chunk policies and isolates the effect of adapting the execution horizon from changes to the underlying action generator. Across our evaluations, DEHP improves the success rate of different high-precision and long-horizon manipulation tasks by a large margin. Our qualitative analysis further shows that DEHP predicts shorter execution horizons during fine-grained stages of the task and longer horizons during free-space motion. In this way, DEHP balances the efficiency of open-loop chunk execution with the reactivity of closed-loop single-step control. Project page: https://dehp-chunking.github.io/

Abstract

Explore similar work

Dynamic Execution Horizon Prediction for Chunk-based Robot Policies

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

Continue or Replan? Bernoulli-Continuation Policy Learning for Adaptive Horizon Execution