Adaptive Action Chunking at Inference-time for Vision-Language-Action Models
Authors: Yuanchang Liang, Xiaobo Wang, Kai Wang, Shuo Wang, Xiaojiang Peng, Haoyu Chen, David Kim Huat Chua, Prahlad Vadakkepat
Organizations: National University of Singapore · Shenzhen University of Advanced Technology · Sangfor Technologies Inc. · Mininglamp Technology · Shenzhen Technology University · City University of Hong Kong
In Vision-Language-Action (VLA) models, action chunking (i.e., executing a sequence of actions without intermediate replanning) is a key technique to improve robotic manipulation abilities. However, a large chunk size reduces the model's responsiveness to new information, while a small one increases the likelihood of mode-jumping, jerky behavior resulting from discontinuities between chunks. Therefore, selecting the optimal chunk size is an urgent demand to balance the model's reactivity and consistency. Unfortunately, a dominant trend in current VLA models is an empirical fixed chunk length at inference-time, hindering their superiority and scalability across diverse manipulation tasks. To address this issue, we propose a novel Adaptive Action Chunking (AAC) strategy, which exploits action entropy as the cue to adaptively determine the chunk size based on current predictions. Extensive experiments on a wide range of simulated and real-world robotic manipulation tasks have demonstrated that our approach substantially improves performance over the state-of-the-art alternatives. The videos and source code are publicly available at https://lance-lot.github.io/adaptive-chunking.github.io/.
Figures & tables
Figure 1: Effects of action chunk sizes . At inference-time, the success rates of the GR00T N1.5 [ 2 ] on different tasks of RoboCasa Kitchen [ 28 ] are highly related to the action chunk size. It can be observed that it is difficult and sub-optimal to empirically set a fixed value for various manipulation tasks.
Figure 2: An overview of AAC . The proposed Adaptive Action Chunking (AAC) algorithm operates solely at inference-time, without any extra training or architectural changes. Specifically, we exploit the action entropy of continuous and discrete values as the cue to adaptively determine the optimal chunk size h∗ in each action chunk at the current observation. Therefore, we can achieve a favorable trade-off between consistency and reactivity in the entire episode, and substantially improve success rates across a variety of manipulation tasks.
RoboCasa
LIBERO
Method
Relo.
Cont.
Rot.
Button
Avg.
Spatial
Long
Object
Goal
Avg.
GR00T (Default h =16)
42.1
80.3
57.6
80.3
59.7
93.6
88.8
97.2
96.8
94.1
GR00T ( h =2)
29.6
57.0
59.4
64.7
47.0
93.6
81.8
93.6
91.6
90.2
GR00T ( h =4)
37.9
71.3
61.2
78.7
56.2
94.0
85.8
95.4
95.2
92.6
GR00T ( h =8)
42.8
80.3
62.6
81.7
61.2
92.0
93.4
97.8
95.4
94.7
GR00T ( h =12)
42.9
79.5
60.0
79.3
60.2
93.0
91.0
97.6
95.2
94.2
Table 1: Main Results on RoboCasa and LIBERO Benchmarks . We report the success rate (%) for various subsets and the overall average. Our AAC achieves the best or competitive results across both benchmarks. Bolded entries indicate the highest success rates.
Figure 3: Rollout of chunk sizes from AAC . The derived chunk sizes align with human intuitions with respect to different semantic phases: a large chunk size is observed during the transportation stage, while a small chunk size appears at the critical manipulation stage.
Figure 4: Distribution of chunk size decisions from AAC . We show the chunk size distribution of episodes on the first task of LIBERO-Spatial: ”Pick up the black bowl next to the cookie box and place it on the plate” . The heatmap indicates the frequency of different chunk sizes at different decision timesteps. The red curve shows the mean chunk size at different observation timesteps.
Method
Spatial
Long
Object
Goal
Avg.
π0.5
98.5
92.5
98.7
98.1
97.0
π0.5 +AAC ( Ours )
99.1
95.2
99.2
98.0
97.9
Table 2: Success rates (%) with π0.5 backbone on LIBERO.
Perturbation level
× 0.2
× 0.3
× 0.4
Avg.
GR00T
11.3
0.4
0.0
3.9
GR00T+AAC ( Ours )
17.9
0.9
0.0
6.3
π0.5
53.2
29.9
9.5
30.9
π0.5 +AAC ( Ours )
57.4
35.3
11.8
34.8
Table 3: Success rates (%) on LIBERO-Pro [ 50 ] .
#Samples
1
5
10
20
30
40
Succ. (%)
94.1
94.7
94.4
95.0
95.0
95.5
Time (ms)
83.0
83.5
84.3
106.0
136.5
157.0
Table 4: Success rates (%) on LIBERO and inference-time under different numbers of samples for estimating action entropy.
Figure 5: Execution examples for real-world tasks using AAC. Videos of complete execution trajectories will be publicly available.
Task
Banana
Button
Drawer
Avg.
GR00T
70.0
65.0
65.0
67.0
GR00T+AAC
90.0
75.0
80.0
82.0
Table 5: Success rates (%) on real-world applications.
Figure 6: AAC improves action accuracy and safety . Left : the gripper collided with the table. Right : the gripper reached an appropriate lowest point.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Task Group
Task Name
GR00T
GR00T + AAC
Relocation
Counter to Cab
57.0
59.0
Counter to Stove
41.0
40.0
Sink to Counter
46.0
52.0
Stove to Counter
60.0
63.0
Counter to Sink
26.0
28.0
Counter to Microwave
33.0
30.0
Appendix
Table 6: Success rates (%) on RoboCasa.
Task Group
Task Name
Fixed-2
Fixed-4
Fixed-8
Fixed-12
Fixed-16
Relocation
Counter to Cab
51.0
68.0
61.0
57.0
57.0
Counter to Stove
31.0
40.0
46.0
45.0
41.0
Sink to Counter
35.0
44.0
56.0
50.0
46.0
Stove to Counter
59.0
62.0
69.0
64.0
60.0
Counter to Sink
6.0
10.0
15.0
19.0
26.0
Counter to Microwave
7.0
12.0
29.0
23.0
33.0
Appendix
Table 7: Success rates (%) with Different Chunk Sizes.