Predict actions from observations.

This is the same reference used in Technical mode. Expand it whenever you need the exact numbers, assumptions or further explanation.

Start with the task

A student driver can match familiar steering demonstrations yet fail after drifting away from the road centre. Robot imitation has the same feedback problem.

Mechanism and units

The network receives six numbers: puck x/y, goal x/y and two instruction indicators. Eight tanh hidden units produce six tanh outputs: three planar commands. Training minimizes mean half-squared command error. Every gradient uses the same preceding checkpoint; the update rate is 0.6 and the interface performs at most five updates per action.

Our coordinates come directly from the rendered scene, so this is learned control from visual positions, not a learned image detector. The teacher is a proportional controller with saturation. Compare it directly with the network: a learned approximation is not automatically better than its teacher. Training only sees supplied and explicitly recorded training trajectories.

Work through the evidence

Compare initial and trained policies at the same start and disturbance. Inspect the observation used for a poor action. Was that observation represented in collection? Add a recovery demonstration before claiming the architecture needs to be larger.

Train the policy, introduce a 0.3 m displacement and inspect its recovery.

What this experiment can establish

Repeated public development runs support debugging, not a final generalization claim. Four hundred updates is a work limit, not a convergence guarantee. A policy checkpoint does not authenticate how it was trained.

Keep collection, fitting, frozen execution and interpretation as separate records. Save the configuration and complete result, including failures, in the robot-learning project. Report submission checks required evidence and text presence; reasoning and implementation quality need independent review.