Quadruped robots that need to work around people and on rough ground face a trade-off. Traditional pipelines separate path planning from low-level locomotion, which can ignore how the robot's whole body interacts with terrain. Fully end-to-end policies avoid that split but demand high-frequency sensing that is often noisy and expensive, and many assume the world is static.
The TRAIN framework, described in an arXiv preprint, uses three reinforcement-learning pipelines. TRAIN-Loco trains an asymmetric actor-critic policy for agile locomotion over uneven terrain without explicit terrain or contact observations. TRAIN-Nav uses a symmetric actor-critic setup to map LiDAR data directly to actions, handling pedestrians under differential-drive kinematics. A unified pipeline combines both.
The authors report benchmarks against locomotion and navigation baselines, and hardware experiments that suggest the policies transfer from simulation to real quadruped robots. The work is a step toward robots that can move through human-populated, uneven environments without separate high-level and low-level controllers.