Real-Time Parameter Control for Trajectory Generation Using Reinforcement Learning With Zero-Shot Sim-to-Real Transfer

Research on trajectory generation algorithms for unmanned ground vehicles (UGVs) has been actively conducted due to the rapid increase in their use across various fields. Trajectory generation for UGVs requires a high level of precision, as various parameters determine the trajectory’s efficiency and safety. Notably, maximum velocity and acceleration are critical factors impacting trajectory performance. In this paper, we propose a novel algorithm that utilizes reinforcement learning to determine the optimal maximum velocity and acceleration in real-time within dynamic environments. The proposed algorithm overcomes the limitations of traditional fixed parameter settings by determining parameters through real-time environmental adaptation. Furthermore, we also propose a PX4-ROS2 based reinforcement learning framework for achieving stable zero-shot sim-to-real transfer. Experimental results in simulation and real-world environments show that the proposed method significantly improves trajectory safety and efficiency while also demonstrating excellent adaptability to changing environments. Furthermore, validation through identical experimental results in both simulation and real-world environments confirms a stable zero-shot sim-to-real transfer.