ACRO: Actor-Critic Rollout Orchestration

Junhoo Lee; Seungyeon Kim; Baekseung Kim; Minkyu Kim; Suhyun Jeon; Jimyeong Kim; Nojun Kwak

cs.ROOct 14, 2026

ACRO: Actor-Critic Rollout Orchestration

Authors: Junhoo Lee, Seungyeon Kim, Baekseung Kim, Minkyu Kim, Suhyun Jeon, Jimyeong Kim, Nojun Kwak

Organizations: Korea Advanced Institute of Science and Technology (KAIST) · Seoul National University

Abstract

Executing a vision-language-action (VLA) policy involves deciding when to continue an attempt and when to intervene to improve its chances of success. We introduce ACRO, an Actor--Critic rollout orchestration layer that makes these decisions for a fixed VLA. ACRO seeks interventions that offer a more promising continuation, are physically feasible, and leave sufficient time to complete the task. A Critic learned from policy rollouts estimates success prospects to decide when to intervene and select a re-entry target along the executed trajectory. Trajectory Retraction shortcuts the visited path to return the robot to that target, where the same VLA resumes execution from a fresh observation. We evaluate ACRO with GR00T N1.7 and $π_{0.5}$ across SimplerEnv, RoboCasa, and real-robot manipulation. Under the same physical execution budget, including all return motion, ACRO improves success by 10.4, 7.3, and 21.7 percentage points, respectively. It also completes additional tasks as the execution budget grows. These results show that controlling when and where a policy resumes can turn its existing capabilities into more successful task executions. Project page: https://junhoo.me/acro

Figures & tables

Panel of Figure 1 — Figure 1: Evaluation environments. (a) Real robot: pick and place, cube stacking, and button pressing. (b) RoboCasa ( Nasiriany et al., 2024 ) : cabinet interaction and pick and place. (c) SimplerEnv Bridge ( Li et al., 2025 ) : eggplant in basket and spoon on towel. Scenes are ordered left to right within each group; simulation images are from official project demonstrations.

			Baselines
Policy	Task	$N$	Base	RoboMonkey †	CycleVLA †	ACRO
GR00T N1.7	Carrot on Plate	60	50.0	31.7	48.3	68.3
SimplerEnv Bridge	Eggplant in Basket	60	36.7	50.0	66.7	51.7
	Spoon on Towel	60	93.3	90.0	93.3	93.3
	Stack Cube	60	18.3	18.3	13.3	26.7
	Overall	240	49.6	47.5	55.4	60.0

Table 1: Task success rate (%) in simulation. All methods run the same base policy under the benchmark’s standard physical step limit, including recovery motion: 300 steps on SimplerEnv Bridge and each task’s registered limit on RoboCasa. Overall pools all trials of a benchmark. Best result per row in bold.

\dagger

Adapted baselines; see Section 4.1 .

Task	$N$	Base	ACRO
Pick and Place	20	70.0	80.0
Stack Cube	20	20.0	55.0
Press Button	20	40.0	60
Overall	60	43.3	65

Table 2: Real-robot success rate (%).

Figure 2: A small rollout set suffices to train the Critic. Success and AUROC of minimum rollout Q over three tasks at

2H_{\mathrm{ref}}

Method	Success (%)	Returns	Steps
Base VLA	60.7	0	0
Periodic	59.5	1.04	92.7
ACRO	68.0	1.06	46.3

Table 3: ACRO improves on scheduled rewinds. All 600 RoboCasa trials at

H

; returns and return steps are trial means.

Category	Gained	Lost	Base	ACRO
	Outcome changes		Mean rollout steps
Open/close	33	4	362.6	296.1
Pick/place	17	14	387.0	345.5
Control	20	8	139.3	157.7
Overall	70	26	327.9	287.3

Table 4: Gained successes exceed losses. All 600 RoboCasa seeds at

H

, relative to Base. Policy-step means use the 46 shared successes after ACRO’s first alarm.

Figure 3: Retraction enables a successful re-approach. ACRO intervenes at step 130, spends 50 steps returning toward the target at step 50, and succeeds at step 224; paired Base VLA times out at 1,000. Q curves show the first 240 steps, with no queries during retraction. Frames replay recorded actions, verified against per-step success flags.

Figure 4: Success grows with budget. Bands: 95% bootstrap intervals.

Path	Success	Steps	Reach
Reverse replay	70.7	62.0	86.2
RDP-simplified	70.9	43.4	80.5
Straight line	70.7	20.6	64.1
Ours	71.3	32.4	85.4

Table 5: Shortcuts reduce return cost. Paths at

2H_{\mathrm{ref}}

: success (%), mean return steps, and target reach (%).

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Category	Task	Limit (steps)
Open/Close/Slide	Close Fridge	1200
	Open Cabinet	1400
	Open Drawer	1000
	Open Stand Mixer Head	600
	Slide Dishwasher Rack	600
Pick-and-Place	Counter to Cabinet	1000

Appendix

Table 6: RoboCasa tasks and their registered episode limits.

Critic
	SimplerEnv	RoboCasa
Base policy	GR00T N1.7	$\pi_{0.5}$
Step budget	300	registered limit ( Table 6 )
Rollouts (train / validation / calibration)	780 / 150 / 300	720 / 180 / 300
Summary network	GRU, width 64	GRU, width 128
History length	8 queries	8 queries

Appendix

Table 7: Hyperparameters of ACRO and the baselines.

Critic	Params	Q AUROC	Success
GRU (ours)	0.44M	0.897	71.3
MLP	0.35M	0.886	71.1
KAN	1.17M	0.939	77.6
Transformer	0.62M	0.915	73.6

Appendix

Table 8: Critic architecture. The MLP and KAN read only the latest query; the GRU and Transformer read the last eight. Success rate (%), averaged over three tasks.

Figure 5: Intervention trigger. Gray points threshold Q at conformal level

\alpha

; both gates are calibrated at

\alpha=0.2

. Failures caught and false alarms are the fractions of executions that fail or succeed under Base VLA in which the trigger intervenes. Averaged over three tasks (450 paired trials).

Figure 6: Success over the execution budget by task. Sixty paired trials per task on SimplerEnv Bridge, comparing Base VLA, RoboMonkey, CycleVLA, and ACRO. All panels use the same success-rate scale. Bands show 95% bootstrap intervals over object placements.

		Policy steps		Physical steps
Category	$N$	Base	ACRO	ACRO
Open/Close/Slide	20	362.6	296.1	365.2
Pick-and-Place	17	387.0	345.5	391.2
Control	9	139.3	157.7	187.0
Overall	46	327.9	287.3	339.9

Appendix

Table 9: Shared successes after ACRO’s first alarm at

H

. Means count every policy attempt; ACRO physical steps also include return motion.

Robotic Manipulation

Robotic manipulation studies how robots perceive, plan, and control physical interactions to grasp, move, assemble, and use objects.

See all 1,699 papers

When to Switch: Reliable Action-Chunk Extension for Vision-Language-Action Models
Oct 5, 2026Seonghoon Yu, Dongwon Kim, HyungRok Jung +4Efficient VLA Model Inference Robotic Manipulation
KAIST · GIST · POSTECH +1
RePO-VLA: Recovery-Driven Policy Optimization for Vision-Language-Action Models
May 10, 2026Weijia Liufu, Xiaoyu Guo, Ruiyi Chen +16VLM Robustness Robot Failure Recovery
Sun Yat-sen University · South China University of Technology · Peng Cheng Laboratory +3
Fast-VLA: You Only Think Once for Robot Manipulation
Oct 15, 2026Kehong Gong, Guanli Hou, Xiaoxu Zheng +6Vision-Language-Action Models Closed-Loop Control of VLA Models
Huawei Technologies Co., Ltd · Huawei Central Research Institute

Oct 5, 2026J/K move · Enter open · S save

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

Seonghoon Yu, Dongwon Kim, HyungRok Jung, Yoonjae Baek+3

KAIST · GIST · POSTECH+1

Efficient VLA Model Inference Robotic Manipulation Vision-Language-Action Models

Open paper AskPDF arXiv

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

Abstract

Robotic Manipulation

Explore similar work

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

RePO-VLA: Recovery-Driven Policy Optimization for Vision-Language-Action Models

Fast-VLA: You Only Think Once for Robot Manipulation