기계공학부 구성원들의 많은 관심과 참여 부탁드립니다.

▣ 주   제: Self-Supervised Robot Motion Learning with Physics Priors

▣ 연   사: Prof. Ahmed Qureshi

▣ 소   속: Purdue University

▣ 일   시: 2025. 9. 26.(금) 16:00

▣ 장   소: 제4공학관 D604호

▣ 초   록

This talk will outline the use of Partial Differential Equation (PDE)-based physics priors for creating efficient plug-and-play algorithms in robot motion learning. These algorithms achieve high efficiency in both training and inference, while remaining effective in complex, high-dimensional environments under diverse constraints.

Recent progress in robot motion learning has largely relied on imitation and offline reinforcement learning, which demand extensive expert trajectories and involve long training times. In contrast, we introduce a new class of self-supervised, physics-informed neural motion policy learners that directly solve the PDEs governing robot motion, eliminating the need for expert data and substantially reducing computational cost.

The talk will also present a novel PDE-derived mapping representation tailored for robot motion generation. Unlike occupancy maps or signed distance fields, this representation is inherently structured for fast and scalable planning. Finally, we demonstrate that these physics-informed approaches surpass state-of-the-art imitation and offline reinforcement learning methods in terms of scalability, efficiency, planning speed, and motion quality, leading to significantly higher overall success rates.