Commit While Futures Agree: Consequence-Aware Adaptive Action Chunking for Robot Manipulation

Yuyan Li; Yujia Wang; Yusong Huang; Junjie Yang; Yanggang Sheng; Ziyi Shi; Wenpeng Xu; Xiaoyang Zhou

cs.ROOct 6, 2026

Commit While Futures Agree: Consequence-Aware Adaptive Action Chunking for Robot Manipulation

Authors: Yuyan Li, Yujia Wang, Yusong Huang, Junjie Yang, Yanggang Sheng, Ziyi Shi, Wenpeng Xu, Xiaoyang Zhou, +3 more

Organizations: The Hong Kong University of Science and Technology (Guangzhou) · ANU Intelligent · Victoria University of Wellington · The Hong Kong University of Science and Technology · Southern University of Science and Technology

Abstract

Action-chunking policies predict multi-step control sequences, but a fundamental question remains: how much of a predicted action chunk should be committed before replanning? Existing systems typically execute a fixed-length prefix, implicitly assuming that the same execution horizon remains trustworthy across states. Some adaptive methods estimate this horizon from the similarity or stability of predicted actions. However, different actions may lead to the same successful outcome, whereas similar actions can produce different futures, suggesting that commitment should be determined by agreement among imagined futures rather than by similarity in action space. To this end, we propose Consequence-Aware Adaptive Action Chunking (CA $^3$ C), an inference-time framework built on a simple principle: commit while imagined futures agree, and replan when they diverge. Without modifying or retraining the base policy, CA $^3$ C uses an action-conditioned world model to imagine the future consequences of multiple candidate action chunks under the same sampling noise. Using these imagined consequences, we formulate execution-horizon estimation as a Bayesian change-point inference problem and select the execution candidate through future consensus. Across multiple simulation benchmarks and real-world robot manipulation tasks, CA $^3$ C consistently improves diverse action-chunking policies, achieving up to a 71.8% relative reduction in failure rate over the corresponding base policies.

Figures & tables

Figure 1: Fixed action chunking executes a preset prefix at every decision step. In contrast, CA 3 C compares the imagined consequences of multiple candidates, commits longer while their futures agree, replans sooner when they diverge, and executes the candidate closest to future consensus.

Figure 2: Overview of Consequence-Aware Adaptive Action Chunking. (a) The diffusion policy samples $K$ candidate action chunks. (b) The action-conditioned world model imagines their visual future states under shared sampling noise. (c) CA 3 C detects where their future latent features begin to disagree, determines the trustworthy horizon $h^{\star}$ , and selects the candidate closest to future consensus. (d) The robot executes the first $h^{\star}$ actions of the selected chunk before replanning.

Method	Spatial	Object	Goal	L-10	Avg.	Spatial	Object	Goal	L-10	Avg.
	ID					OOD
GR00T	99.00	98.00	98.00	95.00	97.50	47.00	50.00	42.00	63.00	50.50
GR00T + CA 3 C (+0.00)	98.00 (-1.00)	98.00 (+0.00)	98.00 (+0.00)	96.00 (+1.00)	97.50 (+0.00)	50.00 (+3.00)	52.00 (+2.00)	42.00 (+0.00)	65.00 (+2.00)	52.25 (+1.75)
$\pi_{0}$	95.00	99.00	97.00	77.00	92.00	48.00	67.00	48.00	61.00	56.00
$\pi_{0}$ + CA 3 C (+0.00)	94.00 (-1.00)	99.00 (+0.00)	95.50 (-1.50)	87.50 (+10.50)	94.00 (+2.00)	48.00 (+0.00)	77.00 (+10.00)	51.00 (+3.00)	65.00 (+4.00)	60.25 (+4.25)
$\pi_{0.5}$	97.00	99.00	96.00	95.00	96.75	74.00	87.00	76.00	78.50	78.88

Table 1: Success rates (%) on LIBERO ID and OOD. Baselines use regular type and rows augmented with CA 3 C are bold . Parentheses show absolute percentage-point changes from each baseline ( red : increase; green : decrease); unchanged values are omitted. The same convention applies to subsequent main-result tables.

Method	Spatial	Object	Goal	L-10	Avg.
GR00T	47.5	48.3	45.8	41.7	45.8
GR00T + CA 3 C (+0.00)	86.7 (+39.2)	80.8 (+32.5)	68.3 (+22.5)	80.0 (+38.3)	79.0 (+33.2)
$\pi_{0}$	75.0	60.8	60.0	75.8	67.9
$\pi_{0}$ + CA 3 C (+0.00)	75.8 (+0.8)	80.8 (+20.0)	60.8 (+0.8)	93.3 (+17.5)	77.7 (+9.8)
$\pi_{0.5}$	93.3	88.3	73.3	97.5	88.1
$\pi_{0.5}$ + CA 3 C (+0.00)	96.7 (+3.4)	96.7 (+8.4)	80.8 (+7.5)	100.0 (+2.5)	93.5 (+5.4)

Table 2: Success rates (%) on 24 LIBERO-Plus tasks. Each suite contains three Robot Initial States tasks and three Objects Layout tasks and is evaluated with 120 rollouts.

Method	Relocation	Container	Rotation	Button	Avg.
GR00T	54.0	80.8	77.0	96.7	70.8
GR00T + CA 3 C (+0.00)	57.5 (+3.5)	85.0 (+4.2)	81.0 (+4.0)	96.7 (+0.00)	74.2 (+3.4)
$\pi_{0}$	19.0	55.8	46.0	56.7	38.5
$\pi_{0}$ + CA 3 C (+0.00)	21.0 (+2.0)	58.3 (+2.5)	49.0 (+3.0)	61.7 (+5.0)	41.3 (+2.8)
$\pi_{0.5}$	44.0	59.2	59.0	60.0	52.9
$\pi_{0.5}$ + CA 3 C (+0.00)	46.0 (+2.0)	63.3 (+4.1)	62.0 (+3.0)	66.7 (+6.7)	56.2 (+3.3)

Table 3: Success rates (%) on the 24 RoboCasa Kitchen tasks. Relocation, Container, Rotation, and Button contain 10, 6, 5, and 3 tasks, respectively; each task is evaluated with 20 rollouts.

Adaptive Horizon	Consensus Selection	Spatial	Object	Goal	L-10	Avg.
		93.3	88.3	73.3	97.5	88.1
	$\checkmark$	95.0	81.7	70.0	98.3	86.3
$\checkmark$		95.0	96.7	80.0	93.3	91.3
$\checkmark$	$\checkmark$	96.7	96.7	80.8	100.0	93.5

Table 4: Component ablation on 24 LIBERO-Plus tasks. The best result in each column is shown in bold , and the second-best result is underlined . Checkmarks indicate enabled components, and values are success rates (%). All variants use the same rollout protocol as the main experiment.

Fixed Horizon	Spatial	Object	Goal	L-10	Avg.
$h=4$	96.7	88.3	78.3	93.3	89.2
$h=5$	90.0	90.0	73.3	95.0	87.1
$h=8$	98.3	96.7	78.3	98.3	92.9
$h=10$	95.0	100.0	80.0	96.7	92.9

Table 5: Fixed-horizon comparison for the

\pi_{0.5}

baseline on LIBERO-Plus (60 rollouts per suite). Best and second-best results are bold and underlined .

Figure 3: Episode-level adaptive commitment for the instruction Put the yellow and white mug in the microwave and close it. Panel (a) traces the selected horizon $h^{\star}$ across the episode and shows representative task stages; the dashed line denotes the fixed $h=5$ baseline, and A/B mark the two decisions enlarged below. Panel (b) shows that disagreement increases after future step 4 at A, so only the trustworthy prefix is executed before replanning. Panel (c) shows consistently low disagreement at B, supporting execution of the full action chunk with $h^{\star}=10$ .

Figure 4: Matched rollouts on LIBERO-Goal OOD for the instruction Open the top drawer and put the bowl inside. The top and bottom rows show $\pi_{0.5}$ and $\pi_{0.5}$ + CA 3 C, respectively, from the same OOD initial state. Columns show representative execution stages, with red and green borders denoting the final failure and success outcomes.

Method	Block Separation	Parcel Sorting	Avg.
ACOT	78.0	56.0	67.0
ACOT + CA 3 C (+0.00)	82.0 (+4.0)	62.0 (+6.0)	72.0 (+5.0)

Table 6: Success rates (%) on real-world tasks using AgiBot G2. Each method is evaluated with 50 rollouts per task.

Figure 5: Suite-level commitment-horizon distributions on LIBERO-Plus. Bars are normalized within each suite; labels show event counts, mean horizons, and frequent-horizon.

Figure 6: Execution examples for real-world tasks using CA 3 C.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: Matched rollout comparison for the instruction Open the top drawer and put the bowl inside.

Figure 8: Matched rollout comparison for the instruction Put the cream cheese in the bowl.

Figure 9: Matched rollout comparison for the instruction Put the wine bottle on top of the cabinet.

Figure 10: Matched rollout comparison for the instruction Put the wine bottle on the rack.

Explore similar work

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
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
ChunkTrust: Adapting Execution Horizons for Robot Policies with Action-Expert Evidence
Sep 30, 2026Fanding Huang, Jingyan Jiang, Shifeng Bao +13Robot Policies Action Chunks
Tsinghua University · Beijing Academy of Artificial Intelligence (BAAI) · Shenzhen Technology University +5

May 30, 2026, cs.ROJ/K move · Enter open · S save

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

Junnan Nie, Jiayi Li, Chenghao Liu, Junyi Lao+4

Peking University · Peking University. · JD Explore Academy.+1

Action Chunks Robot Policies Robot Systems Kinematics

Open paper AskPDF arXiv

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.

Abstract

Explore similar work

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

Dynamic Execution Horizon Prediction for Chunk-based Robot Policies

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