Trajectory training

A trajectory records a sequence of gameplay: the actions taken, the resulting game data and snapshots that can restore the machine later.

The framework splits recorded gameplay into segments around useful events, such as collecting an item or entering a new area. It can retain promising segments and use them to improve later training.

Why keep earlier gameplay?

An agent may discover a useful route long before it can follow that route reliably. Keeping the route lets it practise from useful points and learn from successful actions instead of rediscovering everything from the beginning.

The trajectory store keeps path slices with their start and end information. Configuration controls which segments to retain, how to score them and which to select for practice.

Start partway through a route

Midpoint starts restore a saved state from a retained segment. The agent then plays onward using its current policy.

This gives it practice near a difficult part of a game. The configuration controls how often these starts are used; regular starts still let the tester measure progress on the original task.

Learn from recorded actions

Behaviour cloning trains the policy to reproduce selected actions from recorded gameplay. It adds an imitation loss alongside normal learning.

Selection matters: the framework can favour useful segments rather than treating every recorded action as equally valuable. Cached observations avoid recreating the same screenshots for every learning update.

Scores and selection

Utilities score routes or locations using signals such as progress, novelty or success from that location. Retention decides what remains in the store; selection decides what the learner or actor uses next.

These mechanisms are configured per experiment. The trajectory API describes the store, selection and curriculum configuration types.