PushT
Centered and off-center pushes, recorded from the simulator.
Loading the recorded experiment…
Try an experiment
Same action. Different contact.
Both recordings repeat [0, −0.5]: move upward. In the first, the pusher starts 14 units off center. Contact turns the T. In the second, the centered push reaches the goal after seven actions.
A rollout is simply this sequence: choose an action, advance the environment, record what happened, and repeat. These frames came from the simulator. The buttons replay that evidence.
Live controls require access to my private Tailscale network and the Spark to be online. This recorded explorer works independently.
The Lecture 3 connection
State → action → next state.
The simulator stores position, velocity, orientation, and spin. A picture shows the scene, but a single frame cannot reveal all that motion.
In this SWM checkout, reset() and step() return numeric observation arrays. We also call render() for the image. Open the trace below to compare the observation with the richer mechanical state.
A physics simulator is a world model with specified rules. A learned world model would learn to predict these transitions from data. This lab supplies the environment and the recorded experience for that next step.
State vs. observation, p. 6 ↗
Inside one step, p. 12 ↗
Action as a relative target, p. 16 ↗
Inspect the recorded transition
Mechanical values aid inspection; they are not a complete physics-engine checkpoint. Reproduce using the saved reset settings and action sequence.