What changes after a command?
This experiment learns the next joint positions and gripper opening from current measurements and a requested movement. It compares the learned result with simple predictors on separate simulation groups.
Reading retained evidence…Prediction error
Lower error is better. These are next-boundary actuator measurements, not grasp success or shoe placement. Each transition normally spans 20 ms. Longer trajectories and closed-loop control have not been tested here. Transitions are correlated; the held-out evidence comes from two reset groups and six recordings.
| Predictor | RMSE · mm | Mean absolute · mm | Worst absolute · mm | Scored transitions |
|---|
Where this fits in a skillspace
Human footage supplies visual examples and task structure. Synchronized robot demonstrations supply movement supervision. This pilot connects episode tapes to a compact predictor; it does not train from shoe videos or send commands to a physical arm.
The first shoe task
Move one shoe from a marked pickup area into one marked slot, then verify its position, direction and stable release.
Start with a fixed table and camera, one shoe and tucked laces. The separate collection plan records missing conditions and complete failures. A validated shoe environment, sensed object state, robot adapter and physical trials are still required.
Evidence and limits
Current observations come from native simulator state. The model uses only current joint position, velocity, gripper gap and issued targets; authored phase, future state and final applied controls are excluded from its inputs.