SCOML: Trajectory Planning Based on Self-Correcting Meta-Reinforcement Learning in Hybird Terrain for Mobile Robot

Trajectory planning is important for ground robots to achieve safe and efficient autonomous navigation in unstructured off-road environments. Most existing methods treat each terrain as a single type. However, in the real world, a ground usually consists of hybrid terrains. In this work, we propose a novel trajectory planning network that handles hybrid terrain. To further enhance safety, we have designed a self-correcting structure based on historical planning data. This structure can correct the trajectory when an inappropriate one is planned. To train the network, we introduce a two-stage training scheme based on Offline Meta-Reinforcement Learn- ing, which can train the network with pre-collected non-optimal datasets and reduce the occurrence of hazardous planning. The proposed approach has been evaluated on both simulated datasets and a real robot platform. Compared to state-of-the- art baseline methods, the proposed approach reduces hazardous planning by 59.3% in hybrid terrains.