Worked example: reading the labels in a teleoperated arm episode
示例
A team captures a tabletop pick-and-place task with a teleoperated robot arm. The recorded episode contains: (a) the operator's controller commands at each timestep, (b) the arm's joint positions and velocities, (c) camera frames of the workspace, and (d) a success flag written at the end of the run. The team also runs the same task autonomously and records the same sensor streams plus the policy's chosen actions and the same success flag. Assume both runs succeed.
演示讲解
Step 1 — Identify the decision source in each run. In the teleoperated run, the operator chose each command. In the autonomous run, the policy chose each command. This is the first thing a buyer needs to know, because it determines whether the dataset contains human intent or machine intent.
Step 2 — Identify the action label granularity. The teleoperated episode can be labeled at every timestep with *what the human did*, because the operator's commands are recorded alongside the resulting states. The autonomous episode can be labeled at every timestep with *what the policy did*, but the only signal about whether that was a good choice is the terminal success flag. There is no per-step human correction to learn from.
Step 3 — Check what the success flag adds. The flag is identical in both runs, so it does not distinguish them. It confirms the outcome but says nothing about the quality of the intermediate decisions.
Step 4 — State the resale implication. A buyer training an imitation-style policy needs step-level human action labels, which only the teleoperated run provides. A buyer evaluating a policy's behavior may value the autonomous run for its outcome labels at scale. The two datasets are not substitutes even though the sensor hardware and the task are the same.