Secrets of Walking Robot (NVIDIA Isaac Lab and Isaac Lab Arena)

In this talk

1. 1:02

   Build the robot’s URDF and USD files

   The presenter describes exporting simplified robot parts from Fusion 360 as STL meshes, while using a more precise CAD model for masses and inertias. They explain how to set link and joint origins, calculate centers of mass, convert inertia units from gram millimeters squared to kilogram meters squared, then import the URDF into Isaac Sim and save an instantiable USD asset.

2. 9:29

   Match simulated actuators to the real robot

   The Isaac Lab project separates actuator models, robot assets, and training tasks. The presenter modified an ideal PD actuator model to include acceleration limits and command delay, then adjusted friction, armature, and delay to better match real actuator behavior. Comparisons across hip, knee, and foot joints showed close but imperfect agreement between simulation and hardware.

3. 15:28

   Tune reinforcement learning for transfer

   The policy configuration defines the neural network, while the main task configuration controls terrain, actions, observations, randomizations, rewards, curricula, commands, and terminations. The presenter randomized the robot’s center of mass by up to 5 cm in X and 2 cm in Y and Z, and found that balancing rewards such as action-rate penalties took months. Slower policies transferred better because the real actuators could track their movements more reliably.

4. 21:41

   Improve foot placement and gait symmetry

   To raise the robot’s steps, the presenter added a penalty when foot lift stayed below 5 cm, then added another to limit foot impact speed after early tests showed hard landings. Center-of-mass randomization helped separate the feet, and an additional distance reward reduced feet crossing or getting too close. Isaac Lab symmetry settings encouraged the left and right legs to move equivalently.

5. 26:13

   Explore Isaac Lab Arena and next steps

   The presenter sees Isaac Lab Arena as a developing framework for comparing policies and quickly testing environments, objects, and robots, though its current emphasis is manipulation. They plan to continue improving state estimation, make the robot stand still when it has no movement command, and reduce falls, which still occur roughly every five to ten minutes.