From Simulation to Reality: Swiss Bipedal Robot (+ NVIDIA Jetson raffle)

In this talk

1. 3:21

   Building a simulation that matches the robot

   The robot’s URDF needed accurate mass, center of mass, and inertia values for each part. Skyentific derived these from the CAD design, adjusting material densities to match parts he had 3D-printed and weighed. He also tuned the Isaac Lab actuator model against real actuator encoder data, including acceleration behavior.

2. 8:20

   Getting policy inputs onto the robot

   The walking policy needed joint readings, joystick commands, IMU data, and base velocity. Because estimating linear velocity proved difficult, Skyentific tried alternative neural network configurations and then used a SLAM camera that provided position, orientation, and velocity. He chose a Jetson Orin Nano Super with a Peak M.2 CAN adapter for more consistent control loops than his Raspberry Pi setup.

3. 16:54

   Debugging hardware and sensor failures

   An inverted IMU Z axis made the robot move violently and break both legs; Skyentific repaired them with tougher PLA parts and limited subsequent tests to 10% policy output. He later found over 100 milliseconds of IMU delay caused by an uncleared buffer, and traced falls during sideways and turning motions to the camera’s world reference frame.

4. 21:13

   Progress and remaining work

   After improving the simulation and resolving sensor issues, Skyentific demonstrated the robot moving forward and backward. He described the project as still unfinished and plans a follow-up on Isaac Lab details, while identifying a proper state estimator as a remaining improvement.