Manipulation
Manipulation is where embodied AI is still visibly behind. Language models had a web of text waiting for them; there is no equivalent corpus of a hand closing correctly around an unfamiliar object. Every example has to be performed, recorded and judged.
This is the clearest illustration of why embodied AI is a workforce problem before it is a modelling problem. The field's answer to the data shortage has been to lower the cost of collecting demonstrations — cheaper teleoperation rigs, handheld capture devices, pooled datasets across dozens of labs. Every one of those approaches converts the bottleneck into human hours: someone still has to perform the task, and someone still has to say whether the attempt succeeded.
What this needs from people
- Demonstrations generated on purpose — this data does not exist until someone performs it.
- Frame-level review of contact: where the grasp begins, slips, or recovers.
- Honest success and failure labels, since a near-miss logged as a success teaches the wrong lesson.
- Consistency across thousands of near-identical clips, which is a workforce problem before it is an ML problem.
The research
Where this subsystem actually stands.
Talks & demonstrations
The people building it, in their own words.
Mobile ALOHA — compilation of autonomous skills
ALOHA Robot at Stanford University
Where HSV fits
We are direct about scope here: HSV does not sell a packaged grasp-labelling product. What we bring is the capability underneath it — trained annotation teams, frame-accurate video work, and the double-review system we already run across our LiDAR and evaluation lines. For manipulation programmes, that is the part that is hard to hire for.
Scope a project