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

cs.ROOct 5, 2026

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

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

Organizations: KAIST · GIST · POSTECH · NVIDIA

Abstract

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

Figures & tables

Figure 1: Effect of chunk length

H

on VLABench, with per-episode inference speedup over

\pi_{0.5}

(Appendix E.2 ).

Panel of Figure 2 — Figure 2: Action prediction errors around subskill transitions: (a) an example of transitions between manipulation subskills; (b) action errors spike at the transition point within a chunk, and these spikes grow with the chunk length $H$ ; and (c) RACE reduces these error spikes across chunk lengths $H$ .

Figure 3: Overview of RACE. Given an observation and instruction, the frozen VLM extracts features cached for both passes. RACE predicts a transition-timing prior via an auxiliary one-step denoising pass (Sec. 3.2 ), then restarts denoising from the same noise, conditioned on it (Sec. 3.3 ).

Method	$H_{\text{exec}}$	Spd $\uparrow$	In-dist.		Category		Common.		Instruct.		Texture		Avg.
Method	$H_{\text{exec}}$	Spd $\uparrow$	SR	PS	SR	PS	SR	PS	SR	PS	SR	PS	SR	PS
$\pi_{0.5}$ Baseline	5	$1.00\times$	40.4	56.7	21.4	35.6	17.0	33.7	18.0	35.9	26.0	42.1	24.6	40.8
[0.2pt/1pt] RACE (ours)	5	$0.96\times$	51.6	68.2	25.0	38.1	26.0	40.0	20.8	37.6	29.0	45.3	30.5	45.8
Action-Chunk Extension
[0.2pt/1pt] $\pi_{0.5}$ Fine-tuning	10	$2.01\times$	53.0	67.8	30.2	41.8	24.0	40.5	22.0	40.4	31.8	49.0	32.2	47.9
	15	$2.83\times$	51.4	65.6	26.6	39.6	23.4	40.3	19.0	37.1	29.4	46.3	30.0	45.8

Table 1: Comparison on VLABench. We report the success rate (SR) and the progress score (PS), the fraction of completed sub-tasks. Spd is the per-episode inference speedup relative to the

\pi_{0.5}

baseline (Appendix E.2 ).

H_{\text{exec}}

is the number of actions executed per policy call, and

\dagger

denotes a model fine-tuned to execute that many actions.

\ddagger

indicates the average execution length, since adaptive chunking methods truncate each chunk adaptively. Results for a wider range of chunk lengths are in Appendix B.2 .

Method	$H_{\text{exec}}$	Spd $\uparrow$	PnP	Doors	Drawer	Sink	Stove	Coffee	Micro.	Nav.	Avg.
$\pi_{0.5}$ Baseline	5	$1.00\times$	51.5	57.0	95.0	74.7	40.0	65.3	79.0	22.0	60.4
[0.2pt/1pt] RACE (ours)	5	$0.97\times$	57.2	56.0	93.0	77.3	42.0	68.0	80.0	40.0	63.5
Action-Chunk Extension
[0.2pt/1pt] $\pi_{0.5}$ Fine-tuning	10	1.96 $\times$	62.2	56.0	94.0	75.3	45.0	65.3	89.0	26.0	65.0
	15	2.94 $\times$	60.5	58.0	94.0	68.7	35.0	72.0	91.0	26.0	64.2
	20	3.84 $\times$	56.8	61.0	89.0	62.7	38.0	62.7	87.0	18.0	60.8

Table 2: Comparison on RoboCasa-H50. We report the success rate on eight task categories, with the average weighted by the number of tasks per category. Spd is the per-episode inference speedup relative to the

\pi_{0.5}

baseline.

\ddagger

indicates the average execution length, since adaptive chunking methods truncate each chunk adaptively.

Transition-timing Prediction Head (Sec. 3.2 )	Transition-conditioned Action Generation (Sec. 3.3 )	Training Objectives			Avg. SR
Transition-timing Prediction Head (Sec. 3.2 )	Transition-conditioned Action Generation (Sec. 3.3 )	$\mathcal{L}_{\mathrm{full}}$	$\mathcal{L}_{\mathrm{aux}}$	$\mathcal{L}_{\mathrm{timing}}$	Avg. SR
-	-	$\checkmark$	-	-	28.1
[0.2pt/1pt] -	-	$\checkmark$	$\checkmark$	-	28.4
$\checkmark$	-	$\checkmark$	-	$\checkmark$	30.6
$\checkmark$	$\checkmark$	$\checkmark$	-	$\checkmark$	32.4
[0.2pt/1pt] $\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	33.0

Table 3: Ablation of each component and training objective. The auxiliary one-step pass is always performed when the transition-timing prediction head is present, since the head reads its features;

\mathcal{L}_{\mathrm{aux}}

only controls whether this pass is additionally supervised as a denoiser.

Table 4: Ablation within the transition-timing prediction head (Sec. 3.2 ). Analyses of transition-label validity and sensitivity to PELT settings are provided in Appendices A.5 and C.1 , respectively.

Panel of Figure 4 — Figure 4: Errors at transition and non-transition actions with $H_{\text{exec}}=10$ and $20$ on VLABench, where $\pi_{0.5}$ is fine-tuned to execute $H_{\text{exec}}$ actions per policy call. At each position within a chunk, we compare the error of actions at transition points (transition error) with that of the remaining actions (non-transition error), which we use as a reference for the error growth with prediction horizon.

Panel of Figure 6 — Figure 6: Adjacent action-token similarity.

	Timing prior at inference			Prior shift
	Random	Zero	Predicted (ours)	$-1$	$+1$
Avg.	23.5	31.8	33.0	30.2	30.7

Table 5: Replacement and shift of timing prior.

Method	$H_{\text{exec}}$	SR $\uparrow$	Time $\downarrow$	Idle $\downarrow$
$\pi_{0.5}$ Fine-tuning	5	36%	12.6 s	2.11 s
$\pi_{0.5}$ Fine-tuning	20	48%	11.6 s	0.40 s
[0.2pt/1pt] RACE (ours)	20	66%	11.8 s	0.41 s

Table 6: Results on a real robot.

Appendix figures & tables29 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 7 — Figure 7: Visualization of subskill transitions. Change points detected by PELT ( Killick et al., 2012 ) are overlaid on the action sequences of two VLABench demonstrations, alongside a representative frame for each subskill. Subskill names are generated with Claude Code ( Anthropic, 2025 ) for illustration only; only the transition points are used as labels for policy learning, not the semantics.

Task	# Episodes	Total steps per episode	Transitions per episode
add_condiment	500	$167.6\pm 14.1$	$5.59\pm 0.80$
insert_flower	500	$167.3\pm 15.6$	$5.42\pm 0.87$
select_toy	500	$160.2\pm 13.7$	$4.02\pm 1.09$
select_chemistry_tube	500	$78.4\pm 8.6$	$3.68\pm 0.69$
select_drink	500	$97.4\pm 8.0$	$3.31\pm 0.60$
select_poker	500	$75.4\pm 7.0$	$3.06\pm 0.47$

Appendix

Table 7: Statistics of detected subskill transitions on VLABench training demonstrations. For each task, we report the number of steps and the number of subskill transitions ( i.e. , change points) per episode as mean

\pm

standard deviation.

Figure 8: An example of transition-timing target and prediction on a held-out VLABench demonstration (

H_{\text{exec}}=20

). Top: standardized action signal and PELT change points; the pink-shaded chunk is enlarged in the bottom plot. Bottom: soft label

\mathbf{p}_{t}

and predicted prior

\hat{\mathbf{p}}_{t}

in this chunk.

Figure 9: Qualitative comparison between PELT change points and reference subskill boundaries recorded by the VLABench demonstration generator.

Transition-label sources	Transitions per episode	Precision	Recall
Random	5.0	0.30	0.28
Speed minima ( Nie et al., 2026 )	2.4	0.40	0.21
[0.2pt/1pt] PELT (used)	3.8	0.82	0.70

Appendix

Table 8: Agreement between transition labels and the reference subskill boundaries recorded by the VLABench demonstration generator, on 300 newly generated demonstrations with a matching tolerance of

\pm 3

steps. Precision and recall are counted per transition and pooled over all episodes.

Method	$H_{\text{exec}}$	Spd $\uparrow$	Spatial	Object	Goal	Long	Avg.
$\pi_{0.5}$ Baseline	5	$1.00\times$	98.83	98.17	97.00	93.83	96.96
Action-Chunk Extension
[0.2pt/1pt] $\pi_{0.5}$ Fine-tuning	10	$1.98\times$	97.60	98.80	97.60	94.40	97.10
	15	$2.96\times$	97.40	99.00	97.80	93.60	96.95
	20	$3.91\times$	94.80	97.60	94.00	91.00	94.35
[0.2pt/1pt] PolicyTrim ∗ ( Wang et al., 2026c )	13.75	$4.85\times$	97.80	98.50	98.80	93.30	97.10

Appendix

Table 9: Comparison on LIBERO.

\ast

PolicyTrim trains a separate model for each suite, and its

H_{\text{exec}}

is averaged over the four suites. Its training cost is also substantially higher; see Appendix B.7 . Spd is the per-episode inference speedup relative to the

\pi_{0.5}

baseline.

\ddagger

indicates the average execution length, since adaptive chunking methods truncate each chunk adaptively.

Method	$H_{\text{exec}}$	Spd $\uparrow$	In-dist.		Category		Common.		Instruct.		Texture		Avg.
Method	$H_{\text{exec}}$	Spd $\uparrow$	SR	PS	SR	PS	SR	PS	SR	PS	SR	PS	SR	PS
$\pi_{0.5}$ Baseline	5	$1.00\times$	40.4	56.7	21.4	35.6	17.0	33.7	18.0	35.9	26.0	42.1	24.6	40.8
[0.2pt/1pt] RACE (ours)		5	$0.96\times$	51.6	68.2	25.0	38.1	26.0	40.0	20.8	37.6	29.0	45.3	30.5	45.8
	10	$1.92\times$	57.6	70.9	30.4	41.8	23.2	39.0	25.0	42.0	36.0	52.8	34.4	49.3
	20	$3.68\times$	54.0	68.2	27.2	38.8	28.2	41.4	21.0	39.4	34.8	51.0	33.0	47.8
	30	$5.11\times$	50.2	65.9	24.4	35.9	20.0	36.1	16.4	36.4	31.4	47.0	28.5	44.3

Appendix

Table 10: Results across a wide range of execution chunk lengths on VLABench.

Setting	In-dist.		Category		Common.		Instruct.		Texture		Avg.
Setting	SR	PS	SR	PS	SR	PS	SR	PS	SR	PS	SR	PS
VLM frozen (used)	54.0	68.2	27.2	38.8	28.2	41.4	21.0	39.4	34.8	51.0	33.0	47.8
VLM joint training	57.8	72.1	30.6	43.2	28.6	43.9	31.0	50.4	46.6	62.6	38.9	54.4

Appendix

Table 11: Effect of training the VLM jointly on VLABench with

H_{\text{exec}}=20

. VLM frozen indicates our post-training setting, where the VLM is kept frozen and only the action expert is post-trained. VLM joint training means that the VLM is also updated.

Setting	Training step	Learnable params (M)	Time per update (s)	Wall-clock time (h)	Peak memory (GB/GPU)
VLM frozen (used)	40K	811	3.2	36	18
[0.2pt/1pt] VLM joint training	40K	3,734	8.6	96	40

Appendix

Table 12: Training cost of the VLM-frozen and VLM joint training settings on VLABench with

H_{\text{exec}}=20

. Both models are trained for 40K steps on four RTX A6000 GPUs. Peak memory is the maximum allocated memory per GPU.

Method	$H$	$H_{\text{exec}}$	In-dist.		Category		Common.		Instruct.		Texture		Avg.
Method	$H$	$H_{\text{exec}}$	SR	PS	SR	PS	SR	PS	SR	PS	SR	PS	SR	PS
$\pi_{0.5}$ Baseline	10	5	40.4	56.7	21.4	35.6	17.0	33.7	18.0	35.9	26.0	42.1	24.6	40.8
$\pi_{0.5}$ Fine-tuning	20	20	48.8	65.3	22.8	37.2	20.2	38.2	20.8	39.3	27.8	43.9	28.1	44.8
$\pi_{0.5}$ Fine-tuning	40	20	52.6	68.2	24.0	36.5	22.4	40.5	20.8	40.9	36.0	53.7	31.2	48.0
[0.2pt/1pt] RACE (ours)	20	20	54.0	68.2	27.2	38.8	28.2	41.4	21.0	39.4	34.8	51.0	33.0	47.8
[0.2pt/1pt] RACE (ours)	40	20	56.8	70.2	28.0	38.6	27.4	43.3	28.2	44.9	35.8	51.9	35.2	49.8

Appendix

Table 13: Results with different prediction chunk lengths

H

at the same execution chunk length

H_{\text{exec}}=20

on VLABench, with the

\pi_{0.5}

baseline as reference. Only the first

H_{\text{exec}}

of the

H

predicted actions are executed per policy call.

Table 14: Generalization to other VLA baselines on LIBERO.

Component	$\pi_{0.5}$	ACoT-VLA	RACE (ours)
VLM	53.5 ms	53.5 ms	53.5 ms
Action expert ( i.e. , 10 denoising steps)	30.8 ms	30.8 ms	30.8 ms
[0.2pt/1pt] Explicit action reasoner †	–	32.5 ms	–
Implicit action reasoner †	–	3.8 ms	–
[0.2pt/1pt] Aux. one-step denoising pass (Sec. 3.2 )	–	–	3.0 ms
Transition-timing head (Sec. 3.2 )	–	–	$\sim$ 0.3 ms

Appendix

Table 15: Per-call inference latency, measured on a single NVIDIA RTX A6000 GPU.

\dagger

denotes ACoT-VLA components measured in our setup with its official implementation, where the explicit action reasoner uses an additional 10-step denoising process to generate coarse action trajectories.

Method	Training step	Global batch	Learnable params (M)	Time per update (s)	Wall-clock time (h)	Peak memory (GB/GPU)
$\pi_{0.5}$ Fine-tuning	40K	64	693	2.5	28	17
[0.2pt/1pt] ACoT-VLA † ( Zhong et al., 2026 )	60K	128	1,309	50.6	843	33
[0.2pt/1pt] RACE (ours)	40K	64	811	3.2	36	18

Appendix

Table 16: Training efficiency comparison on VLABench. All methods are measured on four NVIDIA RTX A6000 GPUs, with the compared methods run from their official implementations.

\dagger

: following its official setting, ACoT-VLA freezes the LLM backbone and trains the visual encoder, the action expert, and its components.

Method	Training step	Global batch	Learnable params (M)	Time per update (s)	Wall-clock time (h)	Peak memory (GB/GPU)
$\pi_{0.5}$ Fine-tuning	20K	64	693	2.5	14	17
[0.2pt/1pt] PolicyTrim † ( Wang et al., 2026c )	500 + 500	2,048	693	1,644	457 ( $\times$ 4 = 1,828)	36
ACoT-VLA ‡ ( Zhong et al., 2026 )	20K	128	1,309	49.8	277	33
[0.2pt/1pt] RACE (ours)	20K	64	811	3.0	17	18

Appendix

Table 17: Training efficiency comparison on LIBERO. All methods are measured on four NVIDIA RTX A6000 GPUs, with PolicyTrim and ACoT-VLA run from their official implementations.

\dagger

: PolicyTrim trains a separate model for each of the four LIBERO suites with two-stage reinforcement learning (500 + 500 RL iterations); its values are for a single suite, and the parentheses give the estimated total for all four suites. Its time per update is per RL iteration, including rollout collection.

\ddagger

: following its official setting, ACoT-VLA freezes the LLM backbone and trains the visual encoder, the action expert, and its components.

RACE	$H_{\text{exec}}$	In-dist.		Category		Common.		Instruct.		Texture		Avg.
RACE	$H_{\text{exec}}$	SR	PS	SR	PS	SR	PS	SR	PS	SR	PS	SR	PS
Run 1 (reported)	20	54.0	68.2	27.2	38.8	28.2	41.4	21.0	39.4	34.8	51.0	33.0	47.8
Run 2	20	53.6	68.9	28.8	40.7	29.0	45.0	22.8	40.4	35.0	52.3	33.8	49.5
Run 3	20	52.4	66.8	32.2	42.8	25.8	40.9	23.8	43.1	30.4	49.3	32.9	48.6
[0.2pt/1pt] Mean $\pm$ std	20	53.3 $\pm$ 0.8	68.0 $\pm$ 1.1	29.4 $\pm$ 2.6	40.8 $\pm$ 2.0	27.7 $\pm$ 1.7	42.4 $\pm$ 2.2	22.5 $\pm$ 1.4	41.0 $\pm$ 1.9	33.4 $\pm$ 2.6	50.9 $\pm$ 1.5	33.3 $\pm$ 0.5	48.6 $\pm$ 0.9

Appendix

Table 18: Statistical significance of RACE on VLABench with

H_{\text{exec}}=20

. In addition to the model reported in Tab. 1 (Run 1), we train RACE twice more with different random seeds (Runs 2 and 3). The last row reports the mean and standard deviation over the three runs.

Penalty $\beta$	# of transitions per episode	Avg. SR
2	5.10	33.1
8	1.82	32.8
[0.2pt/1pt] 4 (used)	3.57	33.0

Appendix

Table 19: Sensitivity to the PELT penalty on VLABench with

H_{\text{exec}}=20

. We vary the penalty

\beta

in Eq. ( 8 ) of Appendix A.3 .

Gate $\alpha^{k}$ of Eq. ( 1 ) in Sec. 3.1	Avg. SR
None ( $\alpha^{k}=1$ )	30.7
[0.2pt/1pt] Learnable per step (ours)	33.0

Appendix

Table 20: Effect of the per-step learnable gate on VLABench with

H_{\text{exec}}=20

Transition label	Avg. SR
Hard labels	32.3
[0.2pt/1pt] Soft labels (used)	33.0

Appendix

Table 21: Effect of soft transition labels on VLABench with

H_{\text{exec}}=20

. Hard labels mark only the detected transition step; soft labels spread each transition over neighboring steps with a Gaussian profile (Appendix A.4 ).

Training strategy	Avg. SR
w/o teacher forcing	32.9
w/o jittering	32.8
[0.2pt/1pt] Both (ours)	33.0

Appendix

Table 22: Effect of teacher forcing and jittering on VLABench with

H_{\text{exec}}=20

. The robustness provided by jittering against inference-time transition shifts is evaluated separately in Tab. 24 of Appendix D.2 .

Evaluation type	Metric	Training demos.			Held-out demos.
		$H_{\text{exec}}$			$H_{\text{exec}}$
		10	15	20	10	15	20
Transition-chunk Detection	AUROC	96.0	95.4	94.4	89.2	89.7	86.5
Transition-chunk Detection	Acc.	87.1	87.0	86.0	79.5	79.6	77.0
[0.2pt/1pt] Transition-point Localization	Acc. within $\pm 1$ step	80.3	75.4	72.1	68.2	58.5	50.6
[0.2pt/1pt] Transition-point Localization	Acc. within $\pm 3$ steps	90.6	85.8	83.3	81.0	69.8	65.1

Appendix

Table 23: Accuracy (%) of the transition-timing prediction head on VLABench across execution chunk lengths

H_{\text{exec}}

. Transition-chunk Detection evaluates, for each chunk, whether it contains a transition. Transition-point Localization evaluates, for each chunk containing a transition, where the transition occurs. Multi-transition Localization evaluates how many of the transitions within a chunk are captured, counting every predicted peak rather than only the highest one. Results are reported on both training and held-out demonstrations.

Model	Shift $\delta$ of the transition-timing prior
Model	$-15$	$-10$	$-5$	$-3$	$-1$	$0$	$+1$	$+3$	$+5$	$+10$	$+15$
w/o jittering	28.6 $-$ 4.2	29.8 $-$ 3.0	24.6 $-$ 8.2	24.8 $-$ 8.0	30.1 $-$ 2.7	32.8	30.8 $-$ 2.0	26.4 $-$ 6.4	30.4 $-$ 2.4	28.7 $-$ 4.1	31.1 $-$ 1.7
[0.2pt/1pt] RACE (ours)	32.1 $-$ 0.9	31.0 $-$ 2.0	29.4 $-$ 3.6	30.3 $-$ 2.7	30.2 $-$ 2.8	33.0	30.7 $-$ 2.3	31.2 $-$ 1.8	30.9 $-$ 2.1	30.9 $-$ 2.1	32.1 $-$ 0.9

Appendix

Table 24: Sensitivity to shifts of the predicted transition timing at inference on VLABench with

H_{\text{exec}}=20

. The peaks of the predicted prior are shifted by

\delta

steps within the chunk; peaks shifted beyond the chunk are dropped.

Prior at inference	Training				Held-out
	Switch	Exact $\uparrow$	$\pm 1$ step $\uparrow$	Missed $\downarrow$	Switch	Exact $\uparrow$	$\pm 1$ step $\uparrow$	Missed $\downarrow$
	offset	Exact $\uparrow$	$\pm 1$ step $\uparrow$	Missed $\downarrow$	offset	Exact $\uparrow$	$\pm 1$ step $\uparrow$	Missed $\downarrow$
Shift $k=-3$	$-0.64$	47.3	85.6	2.5	$-0.47$	55.0	82.2	7.5
Shift $k=-2$	$-0.48$	57.9	91.3	1.1	$-0.26$	60.1	89.1	5.1
Shift $k=-1$	$-0.25$	71.9	94.5	1.3	$+0.06$	58.9	90.0	5.6
[0.2pt/1pt] Predicted (ours)	$-0.05$	75.0	97.3	1.0	$+0.29$	53.8	90.8	4.1

Appendix

Table 25: Effect of the transition-timing prior on the generated gripper switches on VLABench with

H_{\text{exec}}=20

. The predicted prior is shifted by

k

steps or replaced with zeros at inference, over chunks where the demonstrated gripper command switches exactly once. Switch offset is the predicted minus the demonstrated switch step. Values under the predicted prior differ slightly from Tab. 27 due to different noise draws.

Error component	MSE		Error reduction (%) over $\pi_{0.5}$ Fine-tuning
Error component	$\pi_{0.5}$ Fine-tuning	RACE (ours)	Error reduction (%) over $\pi_{0.5}$ Fine-tuning
Translation	3.448	2.995	13.1
Rotation	0.277	0.190	31.5
Gripper	0.124	0.094	24.6
[0.2pt/1pt] Overall (7-dim)	1.614	1.378	14.6

Appendix

Table 26: Action errors (MSE) at transitions by action component on VLABench with

H_{\text{exec}}=20

Demos.	Method	Exact $\uparrow$	$\pm 1$ step $\uparrow$	$\pm 3$ steps $\uparrow$	Missed $\downarrow$	Mean error (steps) $\downarrow$	False switch $\downarrow$
Training	$\pi_{0.5}$ Fine-tuning	70.1	94.1	97.4	2.5	0.32	0.5
	RACE (ours)	75.2	97.2	99.0	0.9	0.26	0.6
	Difference	$+5.1$ [3.0, 6.9]	$+3.1$ [2.2, 4.0]	$+1.6$ [1.0, 2.2]	$-1.6$ [ $-2.2$ , $-1.0$ ]	$-0.06$ [ $-0.08$ , $-0.04$ ]	$+0.0$ [ $-0.1$ , 0.2]
Held-out	$\pi_{0.5}$ Fine-tuning	50.6	86.9	93.7	6.3	0.54	0.6
	RACE (ours)	53.8	91.5	95.9	3.9	0.50	0.7
	Difference	$+3.2$ [ $-1.2$ , 7.6]	$+4.6$ [1.9, 7.1]	$+2.2$ [0.5, 4.1]	$-2.4$ [ $-4.3$ , $-0.9$ ]	$-0.04$ [ $-0.09$ , 0.01]	$+0.1$ [ $-0.1$ , 0.3]

Appendix

Table 27: Gripper switching timing on VLABench with

H_{\text{exec}}=20

, over chunks where the demonstrated gripper command switches exactly once, evaluated on 500 randomly sampled training demonstrations and the 300 held-out demonstrations of Appendix A.5 . All values are in % except Mean error , which is in action steps (0.1 s each). Mean error is the mean absolute offset between the predicted and the demonstrated switch, over chunks in which the model predicts a switch; missed switches are counted in the Missed column. False switch is the fraction of chunks without a demonstrated switch in which the prediction switches. Difference is RACE minus fine-tuning on the same chunks, with 95% percentile confidence intervals of this paired difference in brackets, obtained from 1,000 bootstrap resamples of episodes.

	$\pi_{0.5}$		RACE (ours)
	$H_{\text{exec}}=5$	$H_{\text{exec}}=20$	$H_{\text{exec}}=20$
# of policy calls per episode	45	17	17
Actions executed per call	7.0	19.9	20.1
[0.2pt/1pt] Latency per call	271 ms	285 ms	372 ms
Camera capture	122 ms	112 ms	109 ms
Network	49 ms	64 ms	75 ms

Appendix

Table 28: Breakdown of the idle time in Tab. 6 over successful trials. The robot keeps executing the current chunk while the next one is prepared, and pauses only when the latency exceeds this execution buffer, i.e. , the execution time of the actions remaining in a chunk when it arrives, after real-time chunking skips the actions elapsed during inference. Latencies are averaged over calls and, unlike Tab. 15 for latency overhead comparisons, include camera capture, network transfer, and real-time chunking ( Black et al., 2025 ) . Pause per call exceeds Latency

-

buffer because calls within the buffer count as zero pause.

Panel of Figure 10 — Figure 10: Task Visualization. Keyframes of a successful rollout of RACE from the front view.

Model	Env. steps per episode	Policy calls per episode	Spd $\uparrow$
Model	Env. steps per episode	Policy calls per episode	Pooled	Task-balanced	Range across tasks
$\pi_{0.5}$ Baseline	117.9	23.93	$1.00\times$	$1.00\times$	–
[0.2pt/1pt] $\pi_{0.5}$ Fine-tuning	112.3	6.10	$3.92\times$	$3.89\times$	$3.15$ – $4.70\times$
RACE (ours)	113.6	6.16	$3.74\times$	$3.64\times$	$2.54$ – $4.53\times$

Appendix

Table 29: Breakdown of the inference speedup at

H_{\text{exec}}=20

on VLABench, over the 127 episodes that all three models complete successfully. Spd is computed over all episodes pooled and with each task weighted equally; the last column shows its range across tasks.

Episode set	# Episodes	$\pi_{0.5}$ Fine-tuning	RACE (ours)
Common successes with the $\pi_{0.5}$ baseline (used)	228 / 256	$3.82\times$	$3.68\times$
[0.2pt/1pt] Common successes of all three models	127	$3.92\times$	$3.74\times$
with each task weighted equally	127	$3.89\times$	$3.64\times$
All episodes, regardless of success	2,500	$4.00\times$	$3.94\times$

Appendix

Table 30: Inference speedup at

H_{\text{exec}}=20

on VLABench under different sets of episodes. The first row is the setting used in the main comparisons.

	$\pi_{0.5}$	RACE (ours)
Error ratio (transition / matched)	1.34 [1.15, 1.56]	1.30 [1.05, 1.62]

Appendix

Table 31: Ratio of the action error at transitions to that at non-transition positions matched by position within the chunk and by decile of the second-order action difference

c_{h}

, on VLABench with

H_{\text{exec}}=20

. A ratio of 1 means no extra error at transitions. Brackets denote 95% episode-level bootstrap confidence intervals.

Explore similar work

CheckVLA: Execution-Time Verification with Action-Conditioned World Model for Long-Horizon Mobile Manipulation
Jul 29, 2026Yushan Liu, Peibo Sun, Xintao Chao +8Action Chunks Long-Horizon Manipulation
Tsinghua University · Shanghai Jiao Tong University · Peking University +2
Reducing Temporal Redundancy for Efficient Vision-Language-Action Inference
Jul 14, 2026Yuzhou Wu, Yuxin Zheng, Muchun Niu +6Diffusion-Based Vision-Language-Actions Vision-Language-Action Framework
1Tianji KernalMind co ltd · 2Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China · 3Shanghai Jiao Tong University, Shanghai, China +2
SEAM: Smooth Execution of Action-Chunked Motion for Vision-Language-Action Policies
Jul 6, 2026Dijia Zhan, Xuemiao Xu, Jinyi Li +1Flow-Matching Vision-Language-Action Action Chunks
South China University of Technology

Jul 29, 2026, cs.ROJ/K move · Enter open · S save

CheckVLA: Execution-Time Verification with Action-Conditioned World Model for Long-Horizon Mobile Manipulation

Yushan Liu, Peibo Sun, Xintao Chao, Zhenyang Yang+7

Tsinghua University · Shanghai Jiao Tong University · Peking University+2

Action Chunks Long-Horizon Manipulation Confidence Calibration

Open paper AskPDF arXiv

Vision-language-action (VLA) policies commonly execute long-horizon mobile manipulation through open-loop action chunks, issuing multiple actions without receiving new high-level visual input. A committed chunk therefore implies how observations should evolve, but accidental deviations can violate this expectation while the remaining actions continue to propagate the error: commit-time policy confidence cannot react to a deviation that occurs after dispatch, and observation-only anomaly scores lack an action-conditioned reference for separating expected effects from unexplained changes. We propose CheckVLA, which verifies execution with a separately trained, frozen action-conditioned world model. A conformally calibrated risk threshold bounds the episode-level probability of an unnecessary first intervention and determines when to intervene, its exceedance controls how strongly the rewritten suffix retains the superseded chunk, latency-aware hard prefixing restricts replacement to actions that remain deployable, and an event-driven keyframe bank preserves evidence of prior progress across repairs. On RoboCasa365, under a common training recipe and a matched invocation budget, CheckVLA attains a 36.1% average success rate against 27.6% for periodic replanning (+8.5 points). At a matched 5% episode-level false-alarm target, action conditioning raises timely recall to 77.9%, against 48.6% for an observation-only control and 37.9% for an action-shuffled control. These simulation results support action-conditioned verification as a way to restore feedback during chunked execution while keeping the repair consistent with inference latency.

Abstract

Explore similar work

CheckVLA: Execution-Time Verification with Action-Conditioned World Model for Long-Horizon Mobile Manipulation

Reducing Temporal Redundancy for Efficient Vision-Language-Action Inference

SEAM: Smooth Execution of Action-Chunked Motion for Vision-Language-Action Policies