All projectsLegged robotics, robot learning
Quadruped locomotion
Unitree Go2: reinforcement-learned locomotion driven through unknown scenes by LiDAR
Locomotion policies for the Unitree Go2 trained with PPO across thousands of parallel environments and replayed on the full physics model. A speed curriculum takes the policy from a walk to a 3 m/s run over rough terrain, and a LiDAR stack (occupancy map, A*, pure pursuit) drives it through scenes it has never seen.
- Type
- Simulation
- Period
- 2026
- Role
- Training, control and navigation
- Stack
- MuJoCo, Genesis, PPO reinforcement learning, PyTorch, 2D LiDAR mapping, A* planning, Python
- Result
- 30 of 30 navigation goals reached across four unseen scenes, 383 m in total, no falls; 3.8 m/s peak on the park loop.
What I built
- PPO locomotion policies with push recovery, replayed on the full physics model
- Walk-to-run policy, 0.6 to 3 m/s, over stairs, rubble and ridges
- 2D LiDAR occupancy mapping with A* planning and live replanning
The locomotion policy itself
The trained trot policy on the physics model, and the same policy recovering from a sideways shove.