Fall 2021 GRASP SFI - Ashish Kumar, UC Berkeley, “Rapid Motor Adaptation for Legged Robots”

In this talk

1. 1:42

   Train a flexible walking policy in simulation

   Kumar describes training a base policy with reinforcement learning and simulated variations in mass, center of mass, friction, terrain, and motor strength. The policy receives an eight-value representation of these conditions alongside the robot’s state. Training used PPO with one billion samples over 24 hours on a desktop, using rewards for forward motion, energy use, smoothness, and ground impact.

2. 4:32

   Adapt online from the robot’s own motion

   Because real-world conditions are unknown and can change, the method estimates its environment representation during operation from the robot’s movement history. A second training phase teaches an adaptation module to predict that representation from 50 observations, or half a second of history. On the robot, the module ran at 10 hertz while the walking policy ran at 100 hertz, without calibration or synchronization.

3. 7:46

   Test recovery on unfamiliar surfaces and loads

   Kumar shows the robot crossing loose planks, a mattress, foam, slippery plastic, and other household surfaces that were not included in simulation training. It also recovers from slips and added payloads, including an eight-kilogram load on a 12-kilogram robot. Comparisons show stronger performance than the Unitree A1 controller and a version of the method with adaptation frozen.

4. 20:30

   Let energy use shape gaits

   Kumar connects animal locomotion research with a learning setup that minimizes physical energy while targeting different speeds. Without specifying gait patterns, the robot develops walking at low speed, trotting at higher speed, and a bouncing gait with a flight phase at the highest speed. A distilled policy switches smoothly among these behaviors based on requested velocity.

5. 28:39

   Use structured walking when conditions are familiar

   On complex terrain, the robot departs from regular gait patterns when obstacles disrupt its steps. Kumar describes energy efficiency, learned rather than prescribed gaits, and rapid adaptation as design priorities. In discussion, he notes limits including sharp steps, long rises beyond the training range, and sudden downhill friction changes that the robot cannot anticipate.