ReGuide framework improves diffusion policies with self-improving guidance

Researchers have developed ReGuide, a novel framework designed to improve the performance of behavior-cloned diffusion policies, which are often susceptible to covariate shift. ReGuide utilizes Phase-Conditioned Guidance (PCG) to generate corrective rollouts by applying targeted guidance in specific phases of task execution. These successful guided rollouts are then integrated back into the policy through fine-tuning (ReGuide-FT) or retraining from scratch (ReGuide-FS), enabling self-improvement. Experiments on benchmarks like RoboMimic Can, Square, Transport, and Tool Hang demonstrated that ReGuide significantly enhances base-policy success rates, outperforming existing methods in test-time correction scenarios. AI

IMPACT Enhances robustness of diffusion policies against covariate shift, potentially improving real-world robotic task success.

RANK_REASON The cluster contains a research paper detailing a new method for improving diffusion policies. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• Phase-Conditioned Guidance
• ReGuide
• ReGuide-FS
• ReGuide-FT
• RoboMimic
• Square
• Tool Hang
• Transport

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →