Source note
Adaptive Action Chunking at Inference-time for Vision-Language-Action Models
Summary
AAC is an inference-time method for vision-language-action models that picks action chunk size on the fly instead of using one fixed value for a whole episode. It uses action entropy from sampled candidate trajectories to choose shorter chunks when predictions are uncertain and longer chunks when predictions are stable.
Problem
- VLA policies often execute action chunks: several future actions are planned and then run without replanning at every step.
- Fixed chunk size creates a trade-off. Large chunks reduce responsiveness to new observations, while small chunks can cause discontinuities, jerky motion, and more mode-jumping between replans.
- The best chunk size changes across tasks and even across phases of one task, so manual tuning with one fixed value limits performance and scalability.
Approach
- AAC runs only at inference time and does not change training or model architecture.
- The model samples N candidate action chunks in parallel, then estimates uncertainty at each future timestep using entropy: Gaussian differential entropy for continuous translation/rotation actions and standard entropy for the discrete gripper action.
- For each possible execution length h, AAC computes the average action entropy over the first h steps of the predicted chunk.
- It selects the chunk size at the point where the average entropy changes most across chunk lengths, with a lower bound ξ so the robot still executes a minimum number of actions before replanning.
- In practice, high entropy leads to short execution and frequent replanning; low entropy leads to longer execution for smoother behavior and fewer model calls.
Results
- On RoboCasa, GR00T with default fixed chunk size h=16 gets 59.7% average success, while GR00T + AAC reaches 62.0%, a +2.3 point gain across 24 tasks.
- On RoboCasa Rotation tasks, success improves from 57.6% to 61.4% with AAC. On Container tasks, it improves from 80.3% to 82.2%. On Relocation, it improves from 42.1% to 44.4%.
- On LIBERO, GR00T with fixed h=16 gets 94.1% average success, while AAC reaches 95.0%. On the harder LIBERO-Long suite, performance rises from 88.8% to 92.8%, a +4.0 point gain.
- Against other fixed chunk sizes on RoboCasa, AAC's 62.0% average beats h=2: 47.0%, h=4: 56.2%, h=8: 61.2%, h=12: 60.2%, and h=16: 59.7%.
- With a different backbone, pi-0.5 on LIBERO improves from 97.0% average success to 97.9% with AAC; the biggest listed gain is on Long tasks, from 92.5% to 95.2%.
- On LIBERO-Pro OOD perturbation tests, GR00T improves from 3.9% average to 6.3% with AAC, and pi-0.5 improves from 30.9% to 34.8%.