𝚂𝙿𝙸𝙽: distilling 𝚂𝚔𝚒𝚕𝚕-𝚁𝚁𝚃 for long-horizon prehensile and non-prehensile manipulation

Abstract

Current robots struggle with long-horizon manipulation tasks requiring sequences of prehensile and non-prehensile skills, contact-rich interactions, and long-term reasoning. We present SPIN (Skill Planning to INference), a framework that distills a computationally intensive planning algorithm into a policy via imitation learning. We propose Skill-RRT, an extension of RRT that incorporates skill applicability checks and intermediate object pose sampling for solving such long-horizon problems. To chain independently trained skills, we introduce connectors, goal-conditioned policies trained to minimize object disturbance during transitions. High-quality demonstrations are generated with Skill-RRT and distilled through noise-based replay in order to reduce online computation time. The resulting policy, trained entirely in simulation, transfers zero-shot to the real world and achieves over 80\% success across three challenging long-horizon manipulation tasks and outperforms state-of-the-art hierarchical RL and planning methods.