One complete trajectory, laid out for the replay buffer.
Attributes
observationsTensor(T, C, H, W).actionsTensor(T, action_dim) — the action taken into each step, so
actions[t] produced observations[t]. This is the
alignment the world models expect.rewardsTensor(T,).discountsTensor(T,), 0 where the episode terminated and 1 everywhere
else — including where it was merely truncated. Feeds
pcont; ignored by models that hold the discount constant.Notes
The first action is a zero vector: there is no action before the first observation, and the recurrence needs something of the right shape to start from.