From Heightmaps to Cameras: Teacher-Student Reinforcement Learning for Rover Navigation: Master’s thesis
Translated title
From Heightmaps to Cameras: Teacher-Student Reinforcement Learning for Rover Navigation
Author
Sørensen, Thomas Schou
Term
4. semester
Education
Publication year
2025
Submitted on
2025-06-04
Pages
69
Abstract
This thesis investigates a DAgger-based teacher-student framework for transferring navigation behavior from a policy trained on privileged heightmap input to one using noisy RGB-D observations. The goal is to support learning under realistic sensor conditions, where reinforcement learning on RGBD data remains challenging due to its partial and high-dimensional nature. A simulation pipeline was extended to enable separate sensor inputs for teacher and student using the RobuROC4 platform. While the student increasingly aligned with the teacher’s actions, it failed to generalize or perform the task effectively. This is attributed to memorybound dataset handling and limited rollout diversity. The findings suggest that teacherstudent imitation learning holds promise for sensor modality transfer, but depends on scalable infrastructure capable of supporting large and diverse training datasets.
Keywords
Reinforcement learning ; Imitation learning ; teacher-student ; DAgger ; Isaac Sim ; Isaac Lab ; RLRoverLab ; RobuROC4 ; RB Summit ; Leo Rover ; Simulation ; Camera ; RGB-D ; Heightmap
Documents