DREAM-Chunk: Reactive Action Chunking with Latent World Model

2026-06-16

AI Summary

This work addresses the fragility and lack of reactivity inherent in action chunking strategies when executed open-loop under stochastic dynamics, actuation errors, and partial observability. To mitigate these limitations, the authors propose a test-time augmentation method that integrates a lightweight latent world model into the action chunking framework. Without requiring policy fine-tuning, the approach performs predictive rollouts over multiple candidate action chunks and selects the sequence whose predicted states best align with actual observations. This enables dynamic adaptation to perturbations without additional training, substantially enhancing robustness and responsiveness in long-horizon tasks. Experiments on the Kinetix benchmark and four real-world robotic manipulation tasks demonstrate significant improvements in success rates over baseline methods, particularly when demonstration data includes corrective behaviors.

Abstract

Action chunking has become a common interface for vision-language-action (VLA) models, enabling low-frequency policy inference to drive high-frequency robot execution. However, once an action chunk is committed, its open-loop execution can be brittle under stochastic dynamics, hardware execution errors, and partial observability. We propose DREAM-Chunk, a test-time scaling method that augments chunking-based policies with a lightweight latent world model, without requiring additional policy fine-tuning. At test time, DREAM-Chunk samples multiple candidate action chunks, rolls out their predicted latent futures, and selects actions from the chunk whose predicted state best matches the observed rollout. In this way, DREAM-Chunk uses additional test-time computation to cover multiple plausible stochastic futures and improve reactivity during long-horizon chunk execution. On the Kinetix benchmark, DREAM-Chunk improves robustness under increasing action noise and benefits from larger candidate sample sizes, especially when demonstrations contain corrective behaviors. We further validate DREAM-Chunk on four manipulation tasks across two robot platforms and two VLA policies under various sources of stochasticity. Across simulation and hardware experiments, DREAM-Chunk improves the robustness of action-chunking policies in stochastic dynamics.

Problem

• action chunking
• stochastic dynamics
• partial observability
• open-loop execution
• robotic manipulation

Innovation

• action chunking
• latent world model
• test-time scaling
• reactive policy
• vision-language-action