Adaptive Action Chunking via Multi-Chunk Q Value Estimation

2026-05-11

AI Summary

Fixed action chunk lengths struggle to adapt to varying states and tasks, limiting behavioral consistency and value estimation performance in reinforcement learning. This work proposes an Adaptive Chunking Horizon (ACH) mechanism that dynamically adjusts chunk lengths during both training and inference—a first in the field. By leveraging a Transformer-based architecture for joint multi-horizon Q-value estimation, ACH outputs Q-values for all candidate chunk lengths in a single forward pass and selects the optimal length conditioned on the current state. Integrated with an offline-to-online reinforcement learning framework and action chunking strategy, the method significantly outperforms fixed-chunk baselines across 34 complex tasks, demonstrating superior generalization and sample efficiency.

Abstract

Action chunking emerged as a pivotal technique in imitation learning, enabling policies to predict cohesive action sequences rather than single actions. Recently, this approach has expanded to reinforcement learning (RL), enhancing behavioral consistency and reducing bootstrapping errors in value function estimation. However, existing methods rely on a fixed chunk length, creating a performance bottleneck as the optimal length varies across states and tasks. In this paper, we propose Adaptive Action CHunking (ACH), a novel offline-to-online RL algorithm that dynamically modulates chunk length during both training and inference. To find the optimal chunk length for a dynamically varying current state, we simultaneously estimate action-values for all candidate chunk lengths in a single forward pass, using a Transformer-based architecture. Our mechanism allows the agent to select the most effective chunk length adaptively based on the current state. Evaluated on 34 challenging tasks, ACH consistently outperforms fixed-length baselines, demonstrating superior generalization and learning efficiency in complex environments.

Problem

• action chunking
• reinforcement learning
• chunk length
• value estimation
• behavioral consistency

Innovation

• adaptive action chunking
• multi-chunk Q value estimation
• Transformer-based RL
• dynamic chunk length
• offline-to-online reinforcement learning