There are two dominant ways to produce demonstration data for manipulation learning. In teleoperation, a human drives the actual robot through the task, using leader arms, motion controllers, or an exoskeleton, and the robot's own sensors and joint states are recorded. In human demonstration, a person performs the task with their own hands while instrumented cameras record it, typically egocentric head and wrist views plus an external array. The methods are often framed as rivals. They are better understood as instruments with different error bars.
What teleoperation gets right
The decisive advantage of teleoperation is action fidelity. The recorded actions are in the robot's own action space: joint positions, gripper commands, end-effector targets. Nothing has to be inferred or translated. Every demonstrated motion is kinematically feasible on that platform by construction, because the platform just performed it. For behaviour cloning on a specific robot, this is the cleanest data that exists, and observations come from the robot's own cameras, so there is no viewpoint gap at all.
If your platform is fixed, your task is narrow, and your tolerances are tight, teleoperation on that platform is the strongest final-mile data you can buy. We say this without hedging, and we say it as a studio that does not sell robots.
What teleoperation costs
Throughput and money. One robot plus one skilled operator produces one stream of episodes, and the robot is a bottleneck you cannot hire your way around. Operators fatigue quickly under the cognitive load of driving a remote body, and teleoperated motion runs slower and visibly more stilted than natural movement, so you pay premium rates for below-natural speed. The interface itself imposes a dexterity ceiling: bimanual fine motor work through current teleoperation rigs remains genuinely hard, and some contact-rich strategies that human hands find trivially, an operator simply cannot express through a leader arm.
There is also a quieter limitation: the data inherits the platform. Change the gripper, change the arm, and yesterday's corpus is suddenly partially stale.
The labelling difference nobody prices in
One asymmetry rarely makes it into vendor comparisons. Teleoperation gives you some annotations nearly for free: the robot's own state stream marks gripper open and close events, and success can often be checked programmatically against joint and pose data. Human demonstration hands you none of that; phase boundaries and outcomes must be annotated from video by people working to a taxonomy, which adds a real line item to the budget and a real place for inconsistency to creep in. In our own programmes, annotation and its quality control account for a substantial share of delivery effort, and we think buyers should ask any human-demonstration vendor to show their labelling workflow, not just their camera rigs. The raw capture cost per episode flatters human demonstration; the labelled, QC-passed cost per episode is the honest number to compare.
What human demonstration gets right
Scale and naturalness. A person using their own hands works at full speed with full dexterity, which means throughput per hour that teleoperation cannot approach, and contact strategies, regrasp, sliding, bracing with a second finger, that no current interface can express. Capture scales with people and stations rather than with robots. And the strategy aligns with where the field has gone: the large egocentric corpora now used for pretraining, EgoVerse-scale collections among them, are human video, and NVIDIA researchers have reported log-linear gains in policy success as demonstration volume multiplies. Volume on that curve is far cheaper to buy with human hands than with robot time.
What human demonstration costs
The embodiment gap, and it deserves to be stated plainly. A human hand is not a parallel gripper. Recorded human motion contains no ground-truth robot actions, so actions must be recovered downstream: hand pose estimation, retargeting to the robot's morphology, filtering out strategies a gripper cannot execute. This is a real modelling burden, and pretending otherwise is how vendors lose trust. Some fraction of natural human technique will never transfer, and a capture programme has to be designed, task by task, to keep that fraction small: constraining grasps to gripper-compatible pinches where the client needs it, for instance, which is exactly the sort of decision fixed in the task card stage of our capture pipeline.
The middle ground is real
Handheld gripper rigs, portable instruments where a human operates a robot-style gripper carrying its own cameras and pose tracking, have earned their place. They record in gripper action space while keeping human-level throughput, and for tabletop pick-and-place-adjacent tasks they can be the best of both worlds. Their limits appear where the gripper itself does: dense cloth work and fine bimanual coordination lose the most from swapping hands for a tool. We treat them as one more instrument in the kit, not a religion.
How we would actually decide
The rules of thumb we use when a lab asks:
- Platform fixed, task narrow, tolerance tight: teleoperate on your robot, and treat purchased human data as pretraining breadth, not the final mile.
- Coverage across task variation, objects, and environments: human demonstration, because breadth is where its economics are unbeatable.
- Deformables and dense fine motor work: human demonstration, because the teleoperation dexterity ceiling bites precisely here.
- Pretraining a generalist policy: human egocentric video at volume, then a smaller teleoperated set on the target platform for alignment.
That last pattern, broad human corpus plus narrow teleoperated finish, is what most serious programmes converge on, and the two sources cover each other's weaknesses well. Whichever route you take, the sourcing decision matters less than the discipline of whoever executes it; the factors that decide whether episodes are usable at all are the subject of our note on data quality.