Simulator · MuJoCo in the browser
Drive a duck before you get yours.
Same physics (MuJoCo) and the same kind of policy (PPO) you will train. The neural network runs right here in your browser, 50 times a second.
What the network sees
101 numbers per step: gyroscope, accelerometer, your command, position and speed of the 14 joints, the last 3 actions, foot contacts and the gait phase.
What the network decides
14 positions for the joint motors, 50 times a second. No scripted steps: the gait was learned by trial and error over millions of simulated steps.
Now train your own.
If you reserved a Pato you can change the rewards, launch training runs on a cloud GPU and bring the policy back here — and to the robot.
Train my duckModel and pre-trained policy: Open Duck Mini v2, Antoine Pirrone, Apache-2.0.