Waddle Lab
A pugglenaut with a robot's legs. Every step here was recorded from the neural policies that ship on Microduck, a 25 cm bipedal robot — replayed on its exact skeleton.
Waddle Lab
A pugglenaut walking on a real robot’s gait. Nothing runs until you say so.
Live physics runs MuJoCo in your browser and downloads about 4 MB: a 2.1 MB physics engine, 1.4 MB of trained weights and 350 KB of robot model. Nothing is fetched until you press it.
Tap the floor and he walks there. Or drive: W/S forward and back, A/D to turn (arrow keys do the same), Space for the bill. A gamepad works too. There is no sideways: the shipped walking policy cannot strafe, so that command would do nothing.
- gyro
- —
- gravity
- —
- position
- —
- upright
- —
- skill
- idle
Joint motion is recorded from the ONNX policies that ship on Pollen Robotics’ Microduck, replayed on that robot’s exact skeleton, and the ground moves under it at the speed those recordings actually reached — which is well under what you ask for. In playback mode there is no physics, so the pugglenaut cannot fall over — switch to live and it can.
Microduck walks by neural network. Fifty times a second, 61 numbers — its gyro, which way gravity points, where its joints are, what it did last, and what you asked for — go into a policy that answers with 14 joint targets. Those policies were trained in simulation and they ship on the real robot.
To get them here, that control loop was run offline and the resulting joint angles recorded, then blended to follow your command. It is the real gait on the real skeleton, with a baby platypus in a helmet where the duck used to be.
There is no physics in playback, so this pugglenaut cannot be pushed over. Teaching it to fall — and to get back up — means running the simulator itself in the browser, which is the next thing to build.