Train a Unitree Go2 walking policy in Isaac Lab and iterate from a weak baseline to a robust gait.
Instructor: Prof. Ludovic Righetti, NYU Tandon School of Engineering.
Recent quadruped videos demonstrate agile jumping, stair climbing, and dynamic maneuvers that motivate this project’s performance goals for gait quality and robustness.
Many state-of-the-art locomotion policies are trained with deep reinforcement learning such as PPO in massively parallel GPU simulators like NVIDIA Isaac Gym and its modern successor Isaac Lab.
Unitree in the wild: dynamic maneuvers and jumps.
Extreme Parkour with Legged Robots (CMU).
Note: Examples above are external or illustrative videos meant as inspiration; they showcase what robust locomotion policies and systems can achieve when trained and tuned extensively on Isaac Lab.
This is what an IsaacLab RL training looks like in the IsaacSim simulator:
What PPO training looks like in Isaac Lab (not rendered in real time).
The goal is to start from a minimal Isaac Lab training pipeline and significantly improve the learned Go2 walking policy through reward design and robustness techniques.
The provided baseline tracks only linear and angular velocities, which leads to a very poor gait that is unstable and not transferable to hardware without additional shaping and constraints.
This is the starting point with only linear and yaw velocity tracking in the reward, which typically yields unstable contacts, slipping feet, and poor base control that do not transfer to the real robot.
Your task is to improve this policy using principled shaping and robustness strategies.
The clips below illustrate the type of stable, symmetric trot and directional tracking expected by the end of the project, emphasizing clean footfalls, limited slip, and consistent base orientation.
Stable trot at nominal speed with low slip (Hardware).
Example of a good policy trained in simulation.
Treat these as qualitative references for gait quality and document the design choices that lead to similar behavior.
compute_reward_CaT (beware that some reward terms have a weight of 0 and thus are deactivated, check weights in the config file); this implementation includes strong reward shaping, domain randomization, and training disturbances for robust sim‑to‑real, but it is written for legacy IsaacGym and the challenge is to re-implement it in Isaac Lab.