Policy Learning in Geometric Spaces

A full-day workshop of talks, posters, and discussions with the goal to bring experts and experts to be in Policy Learning in Geometric Spaces together from across the globe.

Abstract

The field of robotics is constantly evolving and demands robots to operate effectively in complex, 3D environments. A critical requirement for this is the ability of robots to learn policies for tasks such as grasping, manipulation, motion planning, and more. In recent years, reinforcement learning has emerged as a promising approach for policy learning in 3D geometric spaces. Specifically, researchers have focused on the use of point clouds, neural representations of geometry (NeRFs, neural SDFs, or occupancy networks), occlusion maps, and other techniques to achieve this. The persistent challenge in integrating geometric representation in policy learning lies in effectively bridging the gap between high-dimensional and low-dimensional representations of complex spatial environments. To encourage knowledge sharing and foster collaborations, this workshop aims to bring together researchers and practitioners working in this area to discuss the latest developments and identify challenges. Through this collaborative effort, we hope to further advance the field of robotics and reinforce the importance of policy learning in geometric spaces.