Worked example: classifying a warehouse robot episode
範例
Inputs (hypothetical): A warehouse robot arm performs one pick-and-place cycle. The recording contains 30 camera frames at 10 frames per second, joint encoder readings for six joints sampled at 100 Hz, gripper force readings, and a log of the commanded action sequence (approach, grasp, lift, place). The recording is time-aligned so each frame and reading carries a timestamp. Question: Is this record physical AI data? Justify with the three criteria.
示範講解
Step 1 — Check the embodied source. The data was produced by a physical robot arm operating in a real warehouse. The source is an embodied system, so criterion 1 is satisfied.
Step 2 — Check the sensor stream. The record contains camera frames, joint encoder readings, and gripper force readings. These are measurements of the physical world taken by sensors on the body, so criterion 2 is satisfied.
Step 3 — Check the action or state context. The record includes the commanded action sequence and the joint positions at each moment. This tells us what the system was doing and what state it was in, so criterion 3 is satisfied.
Step 4 — Combine. All three criteria hold simultaneously, and the streams are time-aligned so the sensor readings and actions describe the same moments. The record therefore qualifies as physical AI data.
Assumption: the timestamps are consistent across streams. If the streams were not time-aligned, the record would still be physical AI data in kind, but its usefulness for training or resale would be reduced because the action-to-observation correspondence would be unclear.