Learning robot behaviors via privileged reinforcement learning
Loading...
Date
2025-07-04
Authors
Ντάγκας, Αλέξανδρος
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Legged robots promise unmatched mobility over cluttered or uneven terrain, yet reliable gait control remains an open challenge. Traditional model–based controllers struggle to cope with real–world uncertainties, while recent reinforcementlearning (RL) approaches often rely on privileged or unrealistic terrain information and therefore do not transfer well to hardware. This thesis closes that gap by coupling on-board perception with learning-based locomotion for a quadruped robot.
We first introduce a lightweight, open MuJoCo simulation framework that procedurally generates diverse 3-D terrains and streams realistic sensor data, enabling fast perception-aware training on commodity hardware. A novel LiDAR–inertial pipeline, inspired by Point-LIO, registers point-cloud sequences into a probabilistic 2.5-D grid map. From this grid we extract a high-resolution local heightmap that is fed directly to the policy, giving the RL agent explicit, task-relevant knowledge of nearby geometry without global mapping.
To accelerate learning and improve stability, we introduce our method Phase-Guided Terrain Traversal(PGTT), where we embed phase-guided gait priors into the reward. This yields a curriculum that steers the policy toward coordinated trotting while still allowing adaptation to terrain.
Finally, we transfer the policy to the Unitree GO2 platform. Without finetuning, the robot climbs household staircases and walks over pallets and rubble, confirming the robustness gained from structured heightmap perception.
Overall, this work demonstrates that combining lightweight simulation, LiDAR derived heightmaps, and gait-aware RL produces agile, terrain-adaptive locomotion that bridges the reality gap, advancing the deployment of legged robots in everyday environments. Our method significantly reduces training time and yields a more robust policy that transfers well to real-world environments.
Description
Keywords
Legged robots, Machine learning, Reinforcement learning, Computer vision