Towards Generalizable and Adaptive Multi-Part Robotic Assembly

Yunsheng Tian, Joshua Jacob, Edward Li Gu, Farhad Javid, Pingchuan Ma, Branden Romero, Sachin Chitta, Shinjiro Sueda, Hui Li, Wojciech Matusik

Keywords: Robotic Assembly, Assembly Planning, Contact-Rich Manipulation, Generalization

Abstract: Multi-part assembly presents significant challenges for robotic automation due to the need for long-horizon planning, contact-rich manipulation, and broad generalization capabilities. In this work, we introduce a general dual-arm robotic system that end-to-end assembles multiple parts in simulation without any human effort. For learning contact-rich assembly skills, we propose a simple reinforcement learning framework that generalizes across object geometries, assembly paths, grasp poses, and robotic manipulators by leveraging SE(3)-equivariance. For effective long-horizon planning, our system integrates task-oriented sequence, motion, and grasp planners with a fixture generation method, facilitating multi-step assemblies using a standard dual-arm robotic setup. Following offline planning and training on a standard peg-in-hole benchmark, we perform zero-shot transfer experiments on six unseen multi-part assemblies from different categories. Our system achieves an average success rate of 90% per assembly step with random grasp poses, demonstrating robust performance and adaptability.

Submission Number: 13