ANYmal, a quadruped trained in simulation
ETH Zurich trained a control policy in simulation and transferred it to the ANYmal quadruped: the robot ran at 1.5 m/s against a previous record of 1.2 m/s, and stood up after falling from an arbitrary pose.
Why it matters
A learned policy beat the hand-engineered model-based controller on the same machine, both on speed and on tracking error. And it did something the model-based controller had no procedure for at all: getting up from whatever position the robot had collapsed into.
Jemin Hwangbo, Joonho Lee, Alexey Dosovitskiy, Dario Bellicoso, Vassilios Tsounis, Vladlen Koltun and Marco Hutter made a learned actuator model the key to transfer: it gave the simulation enough accuracy for the policy to cross to hardware. Speed was 1.58 m/s in simulation and 1.5 m/s on the physical robot, a 25% improvement on the previous 1.2 m/s record. The model-based controller’s error proved about 95% higher than the learned controller’s for linear velocity and about 60% higher for yaw rate. Recovery from a fall succeeded 100% of the time once joint velocity constraints were relaxed, and the very first attempt on hardware already worked. Training for locomotion cost nine days of simulated time, which is four hours of computation; training recovery took eleven hours. The simulator ran at nearly 500,000 time steps per second, roughly a thousand times faster than real time. This is the one laboratory in this arc working in Europe.