DREAM-Chunk: Reactive Action Chunking with Latent World Model
TLDR
DREAM-Chunk augments action chunking policies with a latent world model to improve robustness under stochastic dynamics via test-time sampling and selection.
Reasoning
The paper presents a novel test-time scaling method that enhances action chunking without fine-tuning, validated on real robot platforms. However, the abstract lacks details on the world model architecture and training, and the computational cost of candidate sampling is not discussed.
Read-first score
Read-first score 50.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 29.
Field roles
Rank sensitivity
Stability: volatile; rank range: 427.
Keyword Scores
Deep Analysis
Innovations
- Test-time scaling method for action chunking policies using a lightweight latent world model without additional fine-tuning
- Sampling multiple candidate action chunks and selecting based on matching predicted latent futures to observed rollout
- Improving robustness of open-loop action chunking under stochastic dynamics by covering multiple plausible futures
Methodology
DREAM-Chunk augments existing chunking-based policies with a lightweight latent world model. At test time, it samples multiple candidate action chunks, rolls out their predicted latent futures using the world model, and selects actions from the chunk whose predicted state best matches the observed rollout. This uses additional test-time computation to cover multiple stochastic futures without requiring policy fine-tuning.
Key Results
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. It was validated on four manipulation tasks across two robot platforms and two VLA policies under various sources of stochasticity, showing improved robustness of action-chunking policies in stochastic dynamics.