Dynamic Head-on Robot Collision Avoidance Using LSTM

This paper proposes a learning-based algorithm to imitate the head-on obstacle avoidance behavior of humans by the mobile robot. Head-on collision avoidance is the most complex behavior where someone comes directly towards the robot and the robot only gets a limited time to avoid a collision. The robot avoids the dynamic obstacles and leads towards the goal using raw 2D laser sensor readings and goal information respectively. These two behaviors of robots depend on the long-term and short-term memory of the algorithm. To properly address this behavior, we propose a novel architecture of LSTM unit named Navigation LSTM (N-LSTM) that is equipped with greedy gates. Obstacles traveling at different speeds need a different steering mechanism, with larger obstacle speed signifying an urgent steering based on the short-term memory alone. This can be hard to model by a LSTM that takes a spatial information as a raw LiDAR data. The proposed N-LSTM further models unique gates that balance between the short-term behavior of obstacle avoidance and the long-term behavior of goal-seeking based on the relative goal position. The N-LSTM experimentally performs better than different variants of LSTM and two classical approaches of navigation namely dynamic window approach and timed elastic band.