DARE to Navigate: Spatiotemporal Dynamic Augmented LiDAR Encoder for Obstacle Detection and Avoidance in Crowded Environments

Abstract

Navigating in environments with both static obstacles and dense crowds of pedestrians using low-cost sensors and lightweight controllers is crucial for the widespread application of mobile robots, where the key challenge is the localization and tracking of dynamic obstacles using only sensor readings. To address this, a lightweight encoder architecture based only on 2D LiDAR and with a spatiotemporal representation, namely spatiotemporal Dynamic Augmented LiDAR Encoder(DARE), is presented to detect and encode dynamic obstacles for obstacle avoidance. Spatially, we use supervised signals of collision risk in simulation to augment the identification and risk estimation of dynamic obstacles in reality. Temporally, we establish a comprehension of state transition dynamics and implicit dynamic obstacle tracking via predictions of future states and rewards. This architecture can be conveniently integrated into existing reinforcement learning frameworks such as SAC as a plug-in for holistic navigation policy optimization. Extensive experiments across a variety of dynamic scenarios, including real-world deployments, demonstrate our method's robustness and outstanding performance in dynamic and crowded environments compared to state-of-the-art approaches.