Adaptive Action Chunking at Inference-time for Vision-Language-Action Models

1. Problem Statement

Vision-Language-Action (VLA) models, particularly those utilizing diffusion-based action heads, rely on action chunking (executing a sequence of actions before replanning) to reduce compounding errors and improve manipulation success rates. However, current state-of-the-art models (e.g., GR00T N1.5, $\pi_0$) employ a fixed chunk size determined empirically during inference.

This fixed approach presents a fundamental trade-off dilemma:

• Large Chunk Sizes: Improve temporal consistency and efficiency but reduce the model's responsiveness to new information, leading to errors in dynamic environments.
• Small Chunk Sizes: Enhance reactivity and allow for frequent replanning but increase the likelihood of "mode-jumping" (jerky behavior) and discontinuities between chunks.

Existing methods attempt to mitigate this via smoothing (EMA) or search strategies (BID), but they still rely on a fixed global chunk size or require task-specific reward signals (reinforcement learning) that do not generalize to real-world deployment. The core problem is the lack of a mechanism to dynamically adapt the chunk size at inference-time based on the model's current confidence and task requirements without additional training.

2. Methodology: Adaptive Action Chunking (AAC)

The authors propose Adaptive Action Chunking (AAC), a novel inference-time strategy that dynamically selects the optimal chunk size ($h^*$) based on action entropy.

Core Concept

The method operates on the intuition that action entropy reflects the uncertainty of the model's predictions:

• High Entropy: Indicates low confidence/uncertainty. The model should execute a smaller chunk and replan frequently to maintain reactivity.
• Low Entropy: Indicates high confidence/reliability. The model should execute a larger chunk to ensure temporal consistency and efficiency.

Technical Implementation

1. Entropy Calculation:

   • For continuous controls (e.g., translation, rotation), the method calculates Gaussian differential entropy using the covariance matrix of the predicted action distribution.
   • For discrete controls (e.g., gripper open/close), it calculates Shannon entropy based on the probability distribution of the discrete states.
   • The system samples $N$ candidate action chunks in parallel to estimate these distributions.

2. Optimal Chunk Size Selection:

   • The algorithm computes the average action entropy ($\bar{E}_h$) for various potential chunk sizes ($h$).
   • It identifies the optimal chunk size $h^*$ by finding the point of maximum differential increase in average entropy:
     $$h^* = \max(\arg\max_h (\bar{E}_{h+1} - \bar{E}_h), \xi)$$
   • Here, $\xi$ is a lower bound to ensure a minimum action magnitude, preventing the chunk size from becoming too small and compromising efficiency.

3. Execution:

   • Once $h^*$ is determined, the first $h^*$ actions of the predicted sequence are executed.
   • The process repeats at the next inference timestep with new observations.

Key Advantages

• Training-Free: AAC requires no additional training or architectural modifications. It works as a plug-in module for existing diffusion-based VLA models.
• Generalization: It naturally handles both continuous and discrete action spaces and adapts to different robot embodiments (single-arm, bimanual, etc.).

3. Key Contributions

1. Analysis of Fixed Chunking: The paper highlights that fixed chunk sizes are sub-optimal across diverse manipulation tasks, demonstrating strong task-dependency in optimal chunk sizes.
2. AAC Algorithm: Introduction of a simple yet effective algorithm that uses action entropy as a cue to dynamically balance reactivity and consistency at inference-time.
3. Extensive Validation: Comprehensive experiments on both simulated benchmarks (RoboCasa, LIBERO) and real-world robotic tasks, proving superiority over fixed-size baselines and state-of-the-art alternatives.

4. Experimental Results

Simulation Benchmarks

• RoboCasa (24 tasks): AAC achieved an average success rate of 62.0%, outperforming the baseline GR00T (59.7%) and all fixed-size variants (ranging from 47.0% to 61.2%). Notable improvements were seen in "Rotation" tasks (fine-grained control), rising from 57.6% to 61.4%.
• LIBERO (4 suites): AAC achieved a 95.0% average success rate compared to the baseline's 94.1%. The most significant gain (+4%) was observed in the challenging LIBERO-Long suite, which involves long-horizon multi-step tasks.
• Backbone Compatibility: The method successfully generalized to the $\pi_0.5$ backbone, improving its average success rate from 97.0% to 97.9%.
• Out-of-Distribution (OOD): On perturbed tasks (LIBERO-Pro), AAC demonstrated better robustness, improving success rates under high position perturbations compared to fixed baselines.

Real-World Experiments

• Setup: Tested on a Realman single-arm robot with a Mycobot gripper on three tasks: Banana pick-and-place, Emergency button pressing, and a long-horizon drawer task.
• Results: AAC improved the average success rate from 67.0% (Baseline) to 82.0%.
  • Banana Task: Improved grasp alignment.
  • Button Task: Showed superior generalization to unseen button positions.
  • Safety: Visual analysis showed AAC prevented collisions (e.g., with the table) by filtering high-entropy actions, whereas the baseline often collided due to low-confidence predictions.

Efficiency

• While calculating entropy requires parallel sampling, the computational overhead is negligible. With 20 samples, the inference delay is only ~20ms on an A800 GPU, which is acceptable for real-time control.

5. Significance

This work addresses a critical bottleneck in deploying VLA models for real-world robotics: the rigidity of fixed inference parameters. By leveraging action entropy, AAC provides a self-regulating mechanism that mimics human intuition—executing coarse, fast movements when confident and switching to fine-grained, frequent replanning when uncertainty arises.

The method's training-free nature makes it highly scalable and immediately applicable to existing VLA models, significantly enhancing their robustness, safety, and success rates in complex, dynamic manipulation tasks without the need for costly retraining or task-specific reward engineering.