Learning Interaction-Aware Trajectory Predictions for Multi-Robot Motion Planning

Abstract

Multi-robot motion planning without a central coordinator usually relies on the sharing of planned trajectories among the robots via wireless communication in order to achieve predictive collision avoidance. Path planners found in the literature that feature this scheme usually boast levels of performance comparable with their centralized counterparts. However, in practice, communication tends to be unreliable and lead to significant performance degradation, which may eventually result in collisions among robots.

This thesis proposes a data-driven decentralized approach for multi-robot motion planning in dynamic environments. Instead of requiring robots to share their intentions with each other, a model based on recurrent neural networks (RNNs) is used to predict them. This model is trained on data obtained with a centralized sequential model predictive control (MPC)-based motion planner in which intended trajectories of neighboring robots are available while planning. The learned model can be efficiently run online and provide accurate predictions for each robot in the environment based on a horizon of past observations of all robots' states. This model is then incorporated into the MPC planning framework so that it may be run in a decentralized manner. It is finally shown in simulations that the proposed approach achieves a similar level of performance to its centralized counterpart.