The paper presents a two-step method for dynamic obstacle avoidance (DOA) in autonomous vehicles, crucial for any vehicle operating in sea, air, or land. The proposed architecture combines supervised and reinforcement learning. The first step uses a data-driven approach to estimate collision risk with a recurrent neural network. The second step integrates these estimates into an RL agent’s observation space, increasing situational awareness. The method is tested by training various RL agents in an environment with multiple obstacles, resulting in a performance improvement equivalent to halving the number of collisions.
Publication date: 29 Nov 2023
Project Page: n/a
Paper: https://arxiv.org/pdf/2311.16841
