MSTR-PPO: A Multi-Scale Temporal Residual PPO Controller for Freeze-Prone Indoor Navigation

Abstract Local navigation controllers often stop making progress in constrained indoor environments, where narrow geometry and moving obstacles compress the feasible motion space until the robot stalls and falls back on repeated recovery behaviors. We treat such freezing as a failure of local control arising from three observable mechanisms, and pair each with one module: insufficient path horizon with a Multi-Scale Path Transformer, motion ambiguity in single-frame sensing with a First-Order Temporal High-Pass Filter, and unstable learned velocity commands with a Bounded Residual Action, reporting empirical evidence that associates each module with the corresponding condition. The resulting controller, MSTR-PPO, learns only a bounded correction around a Pure Pursuit nominal action, so it drops into the standard sensor–path–velocity-command interface instead of replacing it with an unconstrained end-to-end policy. We compare eight learned local-control configurations alongside three classical baselines across nine Gazebo worlds with dynamic obstacles, using three independent training seeds for each learned controller, and deploy the trained policy as a ROS 2 Nav2 C++ controller. MSTR-PPO clears both freeze-prone worlds with essentially no stack-level recoveries, where the best default-parameter classical baseline does not, and attains 92.1 ± 1.3% average success on the main seven-world suite—the highest of all learned variants—at real-time CPU control rates. Stress tests expose a limitation on continuous-curvature paths, and a preliminary Unitree Go2 test transfers successfully at the same interface.