What actually drives the cost of a robot data collection

A cost model with four lines

Robot collection cost decomposes into four lines, and only one of them is the thing people quote. There is the demonstration time itself, the setup and reset time around it, the fixed engineering work that makes the rig run at all, and the annotation and review that turn episodes into a dataset. Add a fifth line that is not a cost but a divisor: yield, the fraction of recorded episodes that survive quality checks.

The shape of the model matters more than the numbers. Because the fixed line is fixed, cost per usable episode falls steeply with volume, and the first few thousand episodes of a project cost several times what the later ones do. That is why a per-episode price quoted at pilot volume looks alarming and a per-episode price quoted at production volume looks reasonable, without anything about the work having changed.

The ratio that decides most projects

In a manipulation task, the demonstration is short and the reset is not. A pick-and-place episode might occupy five to fifteen seconds of motion, while restoring the scene, restaging the objects, returning the arm to a defined pose, and confirming the setup can take several times that. The ratio of reset to demonstration is the single largest determinant of episodes per shift, and it is set by task and scene design rather than by the operator.

Design the task so that failure is cheap. A task where a missed grasp throws the object across the room costs more per attempt than one where the object falls into a tray, because the difference is a person walking over and searching. Pre-staged object trays, spare objects staged just outside the camera frame, a fixed reset pose, and a written rule for what to do when something lands under the table all reduce this line without touching hardware.

Yield is a multiplier, not a footnote

If forty percent of recorded episodes are rejected for tracking loss, occlusion, aborted attempts, or operator error, then the effective cost per usable episode is roughly 1.7 times the cost per recorded episode. Yield improvements are the cheapest cost reduction available in most projects, because they consume attention rather than capital.

Yield also varies by task, and that variation should feed back into task selection. A task that produces usable episodes at sixty percent costs substantially more per usable episode than one at ninety percent, even when both take the same time to demonstrate, and the difference is invisible in a per-episode rate that counts recorded attempts. Track yield per task and per operator from the first week, because a low-yield task that was chosen for its interesting behavior may not survive the arithmetic.

Where the fixed costs hide

Bring-up is the line most often left out. Integrating an arm with a teleoperation rig, getting joint state at a usable rate, wiring cameras, writing the session recorder, and calibrating everything can take weeks of engineering before the first usable episode exists. That work is real, it is salaried, and it is charged to the project whether or not anyone writes it down.

The rest of the fixed block is maintenance and operations: spare grippers, spare cables, repair time when a crash bends a wrist, storage and transfer for image-heavy episodes, encoding time, and the tooling that runs quality checks. A station that is down costs its whole throughput, not just the repair, which is why stocking the cheap parts matters more than optimizing anything else.

The order to pull the levers in

The levers differ in cost and in how quickly they take effect, so the order is not arbitrary.

• Cut reset time first. It is free, it is immediate, and it usually doubles or triples episodes per shift before anything else changes.
• Raise yield second. Ingest checks and a written session protocol remove the worst rejects without new equipment.
• Increase duty cycle third, if the hardware tolerates it. A second shift on the same rig costs staffing and maintenance, not capital.
• Add stations fourth. A second mid-range rig that produces usable episodes beats one upgraded rig, because throughput is parallel and repair is not.
• Buy better hardware last, and only for a bottleneck that survived the first four steps.

Where cutting cost becomes expensive

Three lines should not be trimmed. Calibration, because uncalibrated episodes carry spatial errors that make them unusable and the defect is only found later. Quality sampling, because unmeasured yield is assumed yield. And spares, because a missing cable converts a cheap day into a lost one.

The comparison that survives all of this is cost per usable episode, counted after quality filtering and after annotation. That number is comparable across rigs, across task choices, and against the alternatives of buying data or generating it in simulation. Cost per recorded episode is not comparable to anything, because it depends on a yield that the quote does not mention.