mono-MORP: Monocular Traversability Representation and Motion Planner for Topological Visual Navigation

Abstract

Topological visual navigation is attractive for biologically-inspired robots because it requires only a monocular camera and can be transferred to robots unseen during training, without dense maps or expensive sensors. Yet many systems rely on waypoint-only open-loop following: even when velocity commands are published at a higher loop rate, the command values are conditioned only on the low-rate waypoint stream and do not incorporate newly observed near-field obstacles. Moreover, waypoint representations predicted from monocular observations tend to bias the controller toward forward motion, making sharp turns or backward maneuvers difficult near obstacles. In this letter, we propose mono-MORP-RL, a reinforcement-learning-based local motion planner that converts monocular depth maps into a metric-free traversability observation by summarizing depth-gradient cues along virtual lanes. Conditioned on this traversability and the predicted waypoint, mono-MORP-RL updates obstacle-aware local commands at the planning rate while using the same waypoint stream, enabling rapid reactions to near-field obstacles without depth sensors or metric-scale depth. In simulations and real-world experiments on a tortoise-inspired mobile robot, mono-MORP-RL increases the success rate and reduces collisions across dynamic obstacle scenarios compared with waypoint-only open-loop followers, and we further analyze the embedded update-rate trade-off against a recent monocular-depth reactive baseline.