ANYmal Learning with Joonho Lee | AI in Robotics Seminar

In this talk

1. 1:53

   Why learn locomotion from data

   Lee explains that conventional controllers simplify robot dynamics to run on limited onboard computing. Those simplifications can leave them vulnerable to slipping, unexpected body contact, and other terrain interactions that are difficult to anticipate and handle with hand-built rules.

2. 9:44

   Training a policy that transfers to the robot

   To avoid risky and costly trial and error on the physical robot, the team trained in simulation. They modeled ANYmal’s actuator behavior using data collected from the robot, then used TRPO to train a neural policy. Their initial terrain randomization alone did not produce reliable walking, in part because the policy lacked useful information about hidden conditions such as foot contact and slope.

3. 18:23

   Learning useful terrain cues and difficulty

   The team fed a temporal convolutional network about two seconds of joint and base measurements so it could infer relevant conditions from movement history, without adding estimated contact states to the deployed policy. A privileged teacher first learned using simulator information such as contact forces and terrain around the feet, then trained the policy through imitation. The team also adjusted terrain difficulty during training based on traversal success, aiming to keep challenges neither too easy nor overwhelming.

4. 26:19

   Results on rough and slippery ground

   Lee reports that a policy trained only on rigid simulated terrain handled varied real-world surfaces, including snow and moving terrain. On a wet whiteboard, ANYmal changed its stepping rhythm after slipping; it also managed added weight and obstacles. The controller ran at 50 Hz, and Lee said it could not anticipate a large obstacle it could not sense.

5. 36:24

   Limits and next steps

   Lee says the approach achieved strong rough-terrain performance without handcrafted contact or slip features. He notes that the team had not proven closed-loop stability and had not found a major benefit from randomizing body mass or dimensions in this study. Follow-up work included adding LiDAR for more active obstacle handling and using collected camera and trajectory data toward visual navigation.