What one environment step returns.
Attributes
observationTensorWhat the agent sees after the step,
(C, H, W) for the pixel
world models.rewardfloatReward for the transition.
terminatedboolThe episode reached a terminal state — a failure, a goal, an
absorbing state. Nothing follows it, so a discount head should
learn
0 here.truncatedboolThe episode was cut short from outside — a time limit, a step
budget. Something would have followed, so the value function
should still bootstrap.
Notes
Termination and truncation are separated because a world model
genuinely needs them apart. Collapsing them into one done flag,
which is the older convention, teaches a discount head that running
out of clock is the same as dying — and then the planner refuses to
look past the horizon it was trained under.