Sample-Efficient Learning-Based Dynamic Environment Navigation With Transferring Experience From Optimization-Based Planner

This letter presents a sample-efficient deep reinforcement learning (DRL) based method to address the intricate task of autonomous navigation in dynamic, constrained environments. The proposed method leverages a graph neural network with feature-wise linear modulation modules to effectively extract features from observations modeled as heterogeneous graph. By transferring offline model predictive control (MPC) experience data, which can be generated fully in parallel, the training process can be jump-started to cope with the sparse rewards. Simulation results demonstrate a 98% navigation success rate, surpassing baselines by at least 6%, while halving the training steps. Furthermore, we successfully deploy our policy network on a real robot and contribute our code to the community through open-sourcing: github.com/TIB-K330/drl_planner.