EASEIR: Efficient and Adaptive Safe-set Estimation via Implicit Representation for High-dimensional Motion Planning

Abstract

Collision-free robotic manipulation is extremely important for all safety-critical applications of robots. Especially for large-scale automation in modern manufacturing facilities where numerous hardware and software systems collaborate in relatively structured environments, accomplishing effectiveness, efficiency, and safety in not only repetitive tasks but also their sporadic reconfigurations is ideal. Yet, existing online and offline Motion Planning (MP) algorithms do not meet such a unique combination of harsh demands, since most of the advances in MP aim for a subset of the requirements. To bridge the gap, we introduce a novel implicit neural function (EASEIR) designed for efficient offline safe set composition for robotic manipulators operating in structured environments. Addressing the challenges of managing high-dimensional configuration spaces (C-space), EASEIR leverages Implicit Neural Representations (INR) to relate coordinates of a discretized robot operation space with collision sets in C-space. EASEIR then utilizes the mapping to actively compose a collision-free set in response to arbitrary occupancy of the operation space by obstacles. The proposed method comprises three core modules: (a) Latent Key Generator (LKG) that maps the coordinates of the space to intermediate latent keys, (b) Latent Key Decoder (LKD) that reconstructs collision sets from the keys, and (c) Full Set Compositor (FSC) that generates a full collision-free set using set operations. On a 6 Degrees of Freedom (DoF) arm, EASEIR generates safe configuration sets nearly 43 times faster than the state-of-the-art analytical method while maintaining comparable accuracy (∼ 0.2% collision) during evaluations in a simulation environment.