What is a robot trajectory in a dataset?
RoboSignal · Published · Version 1.0
In robot datasets, a trajectory usually means an ordered sequence of observations and actions from a robot interaction. An episode may also include task labels, timestamps, rewards or termination information. Always check the dataset’s own counting rule; the word alone does not fix its duration or contents.
The record structure matters
The Open X-Embodiment project provides a concrete reference for robot-learning data across robot platforms. When inspecting any release, follow its loader and schema to see what one episode contains. Check action representation, sampling rate, robot identity and sequence boundaries before combining data from different systems.
Counts do not convert themselves into hours
A trajectory count cannot be converted to duration without measured episode lengths or a stated duration basis. Successful and failed trajectories may differ in length. Truncated records and overlapping camera views can also affect totals. If a source publishes only a count, preserve that count. Do not invent an average episode length to make a scale comparison.
Compare like with like
Distinguish robot-action trajectories from human egocentric video and synthetic sequences. Keep unique elapsed time separate from the sum of simultaneous sensor recordings. Three cameras recording the same one-hour session create three camera-hours and one elapsed hour under full overlap. Our dataset scale calculator shows the arithmetic and its assumptions. It does not inspect or certify a dataset.
Primary references
Reference links checked 2026-10-04. Project claims remain attributed to their original source. This page is not a certification or a live test of the referenced system.
Related reading
Cite this reference
RoboSignal. “What is a robot trajectory in a dataset?” (2026-10-04), version 1.0. https://robosignal.ai/glossary/robot-trajectory. Cite the original project separately for its own reported results.