Demo
CartPole-v1, running in your browser. The physics are the Gymnasium equations rewritten in TypeScript, and the action comes from the trained network replayed from its exported weights. Nothing is computed on a server.
No run has exported weights yet, so this is a random agent. The trained policy will appear here after the first run finishes and export_policy writes its policy.json.
Action probabilities
The random agent has no policy to read. Turn it off to watch the network decide.
- cart position
- 0.000
- cart velocity
- 0.000
- pole angle
- 0.000
- pole velocity
- 0.000
What you are looking at
The agent sees four numbers: the position and velocity of the cart, and the angle and angular velocity of the pole. It chooses to push left or right. The episode ends when the pole passes twelve degrees from vertical, when the cart reaches the edge of the track, or at 500 steps, which is the highest score the environment allows.