
Teleoperation Data Collection: The Bottleneck Isn't the Robot
Collecting data via real-robot teleoperation — the most expensive thing is the operator's time. The robot body itself isn't the big cost.
From what I've seen using this, the most direct impression is: to collect one usable trajectory, only about half of the operator's time is actually spent executing the task — the rest is all wasted on resetting, aligning, and redoing. One hand twitch and the trajectory is ruined. This has little to do with the precision of the teleoperation equipment; the key is whether there's a process that rolls back on error.
I've seen a team doing full-body VR teleoperation, collecting humanoid full-body data and then training a vision foundation policy that predicts full-body joint positions directly from visual observations — the approach is sound. The numbers another platform published are quite telling: across five different controllers they collected about 51.5 hours of simulation data, with only about 9 hours of real-robot data. Simulation to real is pulled to five to one because real-robot collection is too expensive and too slow — anything that can be moved to simulation has been moved there.
So the core competitiveness of real-robot teleoperation data collection at this stage lies in how engineered the collection pipeline is, not in the robot body. Whoever can cut down the grunt work of resetting, calibration, and re-collecting after failures will get more effective data per hour. Just piling up more robots is useless.
Physix Frontier