Coordinated Multi-Agent Motion Planning via Imitation Learning

Abstract

As autonomous vehicles (AVs) become increasingly adopted, many opportunities to share information and communicate with one another will arise. This ability to communicate reduces future uncertainties and allows for collaboration in downstream tasks, such as planning, ensuring increased road- side safety. Currently, many motion planners treat other AVs as standalone human agents and plan to avoid their futures without exploring the potential for collaboration and better- informed motion planning. Towards this end, we present a method for coordinated multi-agent motion planning between two or more AVs that search over a distribution of expert future trajectories to jointly plan paths. We evaluate our model on a didactic, illustrative dataset to experimentally verify its performance, with future plans to use more realistic perception data.