Simulation can now manufacture the training runs a humanoid learns from, which pushes the scarce input upstream to the human motion it imitates.
Teaching a humanoid to open a drawer usually starts with a person in a harness, moving the machine through the motion while the recording runs. It works, and it is also slow, tiring and expensive, and every new task starts the whole process over again. The question the field is chewing on right now is how much of that recording can be handed to a simulator instead.
Rehearsing Without A Rig
A recent preprint called DreamVLA takes a direct swing at that question. The authors propose generating human motion trajectories with a model, then using those generated motions as references to steer reinforcement learning inside a simulator, so a rough imagined movement is converted into one that actually finishes the task. Those converted trajectories become the training set for a vision-language-action policy, the kind of model that takes a camera feed plus a plain-language instruction and produces robot movement.
The piece they describe as their central contribution is the correction loop. When the trained student fails partway through a run, they restart the simulated expert from the point of failure with the same reference motion ahead of it and keep the corrective runs that succeed, which is their answer to the gap between copying a recording and acting in a world that has drifted away from it. They report that the procedure trains policies for tasks such as picking from a platform, opening a drawer, pressing a button, and walking and then picking, with no new task-specific teleoperation recordings.
I think the part worth sitting with is not the saved recording hours. It is the admission buried in that correction loop: imitation of a fixed set of trajectories leaves a hole, and something has to produce the data that fills it. In simulation that something is a reward-driven expert, and it works because a drawer in a simulator is made of numbers somebody already chose.
The Part That Still Has To Be Observed
Generated motion is not free of people either. A model that proposes plausible human movement does so because it was fitted to recorded human movement, so the range of what it can suggest is roughly the range of what was captured in the first place. Ask it for a motion nobody ever recorded and you get a confident approximation. That is tolerable when the simulator can test the approximation and throw out the failures, though the sorting is only as good as the physics the simulator was handed.
Everything the simulator was handed began as an observation of the real world: how much a drawer weighs, how a button resists a finger, how a body shifts its balance while stepping and reaching at the same time. Capturing those things and turning them into records a model can train on is ordinary, demanding operational work, and it sits underneath the simulated pipeline and the real one alike. We develop and deliver embodied intelligence services spanning high-quality data collection, supervised annotation, and AI training data delivery for frontier AI labs, humanoid robotics manufacturers, and autonomous mobility clients worldwide.[1]
What I Would Watch From Here
Contact-rich work is where simulated physics gets expensive and unreliable, because friction, slip and deformation are approximated at best. I would watch whether methods of this kind hold their quality as the tasks move away from rigid objects and clean geometry. I would also watch the reference corpora themselves, since a generator of human motion is a compressed version of whatever library it was fitted to, and the blind spots in that library travel quietly into every trajectory it produces.
The economics strike me as the most underrated part of all this. Simulation keeps driving down the cost of the hundredth variation of a task the field already knows how to describe, and it does almost nothing for the first honest record of a task nobody has captured yet. Those two costs move in opposite directions, and the second one is the one that decides which robots can be built at all.
Every advance that makes robot training data cheaper raises the worth of the few records that cannot be manufactured. Simulators repeat what they have been given, and generators recombine it, which is useful and real and also bounded by the same source. The ground truth underneath both is still gathered by people, carefully, in the physical world, and that is where the binding constraint on embodied intelligence actually sits.
Sources
- (exhibit EX-99.2: PRESS RELEASE), 8-K, filed 2026-08-17, https://www.sec.gov/Archives/edgar/data/1976663/000147793226005111/nexscient_8k.htm
For disclosures, see this article on our website: https://nexscient.ai/the-demonstrations-nobody-recorded/#disclosure
Top comments (0)