Exploiting LiDAR Symmetries to Enhance Value Estimation Policies

Abstract

Motion planning problems frequently exhibit geometric symmetry in both action and dynamics. This symmetric property can be exploited in learning-based solutions such as reinforcement learning during inference time, which allows the learned solution to improve without the need for additional training time. However, existing methods rely on manipulating the explicit state features in the observation such as joint angles, positions, and velocity in feature vectors. These approaches are not feasible for end-to-end solutions that consider raw sensor data such as distance measures from LiDAR sensors. This work proposes to address this limitation by exploiting the symmetric properties of observation itself. By taking advantage of the radial structure of LiDAR beams, geometric rotations can be performed to produce alternate symmetric states efficiently, which in turn are used in a post-hoc symmetry exploitation policy. The rotational nature of this symmetry enables the exploitation of multiple symmetric observations simultaneously. Empirical results show the effect of the amount of observations on the performance benefit and its impact on the processing time. The experiments are performed on a testbed of motion planning tasks of varying difficulty, observation structures, and agent counts. Additionally, hardware demonstrations confirm that the symmetry of simulated LiDAR can be applied to robotic sensor hardware and can perform a real-world task using models trained from simulation. The results show that the post-hoc multi-symmetry exploitation of LiDAR observations improves performance at no training time cost and can be used to reduce training time through early stopping.