A Unified Conditional Policy for Multi-Robot Navigation via LiDAR-to-Vision Distillation

Mobile-robot navigation policies have typically assumed a fixed sensing input and robot platform. In this work, we investigate a teacher–student policy where teachers learn continuous velocity commands from LiDAR-based Twin Delayed Deep Deterministic Policy Gradient, and the student navigates using either LiDAR or camera observations. The student has two modalities with separate feature extractors. The extracted features and robot-state variables are fed into a shared encoder and actor, which map them to linear and angular velocity commands. The generalized two-platform configuration includes a Robot-ID token of scalar type. Training and evaluation were performed in ROS-Gazebo using simulated Pioneer 3-DX and TurtleBot3 Waffle robots. Robot-ID conditioning increased generalized LiDAR success to 90.4% on Pioneer 3-DX and to 91.2% on TurtleBot3 Waffle, up from 71.8% and 79.5%, and camera-based success was 87.3% and 84.0%, respectively. Across four controlled LiDAR-to-vision handover conditions, we achieved an overall success rate of 88.8-90.8%. Overall, 85.2-90.1% of the episodes that were still active at the planned switch were completed without resetting or retraining the policy. The results demonstrate that the conditioning on robot identity is very helpful for cross-platform LiDAR performance and that the shared student policy is able to continue navigation after external scheduling of the active sensing branch. The results provide a controlled, simulation-based feasibility assessment of the proposed policy architecture and establish the basis for subsequent physical robot validation.