Learning to Tune a Mobile Robot Planner: Hierarchical Architecture and Sim‐to‐Real Transfer

ABSTRACT Autonomous navigation in complex, unstructured environments poses a significant challenge, with traditional planners lacking adaptability and end‐to‐end learning methods hindered by data dependency or training instability. Learning‐based hybrid methods that automatically tune traditional planner parameters are a promising compromise, but we identify an underlying architectural limitation that contributes to constraining their performance: a low‐frequency, high‐latency control loop. This paper resolves this architectural bottleneck by proposing a multi‐rate, hybrid hierarchical architecture. Our framework explicitly decouples the task into three coordinated loops: a low‐frequency learning‐based tuning loop, a mid‐frequency model‐based planning loop, and a high‐frequency learning‐based control loop. To train the two learning agents, we introduce cyclic co‐training, a structured curriculum that resolves training instability and allows the tuner to learn more aggressive, high‐performance policies. To ensure robust real‐world deployment, we further design a Terrain‐Adaptive Controller that infers latent environment dynamics from its interaction history, bridging the sim‐to‐real gap by ensuring consistent tracking performance. Extensive experiments demonstrate the effectiveness of our approach. On the BARN benchmark, our method achieves a remarkable score of 0.485 out of 0.500, with an 51.6% improvement over the baseline parameter tuning method. Furthermore, on physical uneven terrains, our approach enhances navigation performance by 58.5% and 38.4% compared to the baseline auto‐tuning methods.