Acrobot playground
Choose two or three links. Drag a joint, press release, and watch a learned controller try to balance it. Each elbow has a motor; the base is unpowered.
M marks each motor. Arrows show torque direction and strength.
−1 ↻ clockwise · counterclockwise ↺ +1
About the models
Two links · one motor
A radial-basis network blends 467 local affine units into one torque command (±1 N·m). Its local units were fitted to nearby examples from a successful swing-up trajectory, labeled by trajectory-tracking and LQR balancing teachers. The inputs are joint angles and velocities.
Three links · two motors
One unified network has three hidden layers of 512 neurons with tanh activations (531,458 parameters). It reads the three angles and velocities and outputs two torque commands (±4 N·m each).
Training combined imitation of the earlier controller with successful optimized recovery trajectories. DAgger added teacher corrections at states the student visited, while rehearsal preserved balancing. The browser runs only this network: no expert switching or fallback controller.
Both models choose an action every 50 milliseconds. Physics and inference run entirely in your browser. Some starting positions may not recover. The three-link weights download is 2.1 MB, loaded only when selected.