Feedback World Model Enables Precise Guidance of Diffusion Policy
TLDR
A feedback world model that corrects predictions online using observed states, improving robotic policy performance under distribution shift.
Reasoning
The paper introduces a novel method to address world model unreliability by closing the loop between prediction and observation at inference time, with theoretical guarantees and real-world experiments. Strengths include practical online correction and convergence guarantees; weaknesses include limited scope to manipulation tasks and lack of explicit comparison baselines in the abstract.
Read-first score
Read-first score 52.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 40.
Field roles
Rank sensitivity
Stability: volatile; rank range: 275.
Keyword Scores
Deep Analysis
Innovations
- Feedback world model paradigm that closes the loop between prediction and observation at inference time
- Lightweight feedback state updated online to iteratively correct future predictions without additional training data or parameter updates
- Action-aware guidance to emphasize action-controllable components while suppressing irrelevant variations
Methodology
The method maintains a lightweight feedback state that is updated online to iteratively correct future predictions, compensating for model errors using real-time observations. It can be interpreted as a latent-space observer with convergence guarantees under mild conditions. Action-aware guidance is introduced to translate corrected predictions into control by emphasizing action-controllable components.
Key Results
The method reduces world model prediction error by up to 76.4% and improves out-of-distribution success rate by 30% on LIBERO-Plus, Robomimic, and real-world manipulation tasks.