Learning spatial representations for efficient robot navigation
OA Version
Citation
Abstract
How can a robot learn to understand and navigate its 3D environment? This research proposes a method for robots to comprehend and traverse their 3D surroundings along a specific route without relying on large architectures or costly 3D sensors like LiDAR, which are impractical for autonomous robots due to computational limitations and excessive power consumption. Instead, it suggests rectifying navigation commands predicted by smaller Imitation Learning models using Reinforcement Learning. The research examines various navigation reward functions used by other studies and introduces a novel multiplicative reward that outperforms them in guiding a robot from a designated starting point through a crowded path to its destination. It also explores motion prediction algorithms based on past motion sequences to anticipate the movements of pedestrians around the robot for facilitating the reinforcement learning algorithm. Moreover, it introduces a custom-designed CARLA RL environment for robot navigation, beneficial for future research in the field. Through experimentation, the research demonstrates that RL effectively corrects and predicts precise actions for the robot to reach its destination, eliminating the necessity for deploying large models for optimal performance.
Description
2024