How learning works
What Happens to a Training Sequence at an Episode’s Edge?
A worked robomimic example shows how sequence padding differs from recorded timesteps and from the earlier context supplied by frame stacking.
A sampled sequence is the window a loader returns for one index, while frame history is earlier context prepended to that sample. Near the end of an episode, an enabled padding rule can fill positions beyond the boundary by repeating the final recorded frame, rather than pulling steps from the next episode.
The official robomimic documentation gives a concrete example. For a demonstration of length 10, with seq_length=10 and access at index 5, the requested window extends beyond the recorded endpoint. With pad_seq_length=True, SequenceDataset repeats the last frame five times to complete the sample. Those copies are padding; they are not five newly recorded robot steps or actions. The documented option controls whether such beyond-boundary subsequences may be fetched.
Frame history handles the opposite direction. In robomimic, frame_stack prepends earlier context; when that history would reach before the first timestep, pad_frame_stack=True permits repetition of the first frame. LeRobotDataset documents temporal windows through delta_timestamps, in seconds relative to t.