GO2 Backflip Training
Role
RL Engineer
Year
2025
Category
Robotics
Context
ETHRC
01.
The Problem
Developing robust locomotion policies for quadruped robots is challenging due to the complex dynamics and the reality gap between simulation and the physical world. The goal was to train a Unitree GO2 robot to perform a backflip, a highly dynamic maneuver.
Leveraging the IsaacLab and Unitree-rl-lab repositories, I trained a PPO (Proximal Policy Optimization) agent in the IsaacSim environment. The training involved curriculum learning and massive parallelization to explore the state space efficiently.
02.
The Approach
03. The Result
The policy was successfully verified using sim-to-sim transfer to Mujoco, demonstrating stable landing and consistent performance. The project highlighted the effectiveness of modern RL pipelines for dynamic robot control.