PEARS: Physical-Prior-Guided Efficient Adaptation via Failure Reasoning and Diffusion Steering for Tactile Manipulation

TL;DR

PEARS is a hybrid reinforcement learning framework that enables efficient real-world adaptation of pretrained robotic policies using tactile feedback. It combines physics-based failure diagnosis with diffusion policy steering to reduce reliance on costly environment interactions. The method achieves high success rates with fewer episodes compared to existing approaches.

Why It Matters

Robotic manipulation in real-world settings often faces unexpected conditions that degrade performance. PEARS reduces the need for extensive trial-and-error learning by leveraging physical priors and efficient policy correction, making deployment faster and more practical.

Key Findings

• PEARS improves success rates by 12.4% to 37.4% over strong baselines in simulation.
• It reduces the number of required interaction episodes by up to 53.2% compared to the fastest baseline.