ALAN-P: Dynamic action pruning for efficient navigation in complex environments

Multi-agent navigation consists on efficiently moving a set of agents from start to goal locations. This is a challenging task since most environments contain static and dynamic obstacles that can significantly restrict the movement of an agent. While existing methods, such as ALAN, can overcome some of these limitations by increasing the action space of the agents, their behavior can be suboptimal in many of the situations that agents can find themselves into. In this work, we propose ALAN-P, a multi agent local navigation method based on the ALAN framework, which improves the agent's behavior by dynamically adapting the action space to the agent's local conditions. The results of our experiments show that the proposed ALAN-P can lead to significant performance improvements over ALAN in a variety of challenging environments.