Collision-Free Robot Navigation in Crowded Environments using Learning based Convex Model Predictive Control

The advent of deep reinforcement learning (DRL) has significantly expanded the application range for autonomous robots. However, safe navigation in crowded and complex environments remains a persistent challenge. This study proposes a robot navigation strategy that utilizes DRL, conceptualizing the observation as the convex static obstacle-free region, a departure from traditional reliance on raw sensor inputs. The novelty of this work is threefold: (1) Formulating an action space that includes both short-term and long-term reference points, based on the robot's kinematic limits and the convex region computed from 2D LiDAR sensor data. (2) Exploring a hybrid solution that combines DRL with Model Predictive Control (MPC). (3) Designing a customized state space and reward function based on the static obstacle-free region, reference points, and the trajectory optimized by MPC. The effectiveness of these improvements has been confirmed through experimental results, demonstrating improved navigation performance in crowded and complex environments.