Analysts expect the field to jump ship to human data for generalization.
Development
A humanoid learns dexterous work from human video with no robot in the loop
Claim
Nvidia trained a humanoid with dexterous hands to assemble cars, operate syringes and fold shirts entirely from more than 20,000 hours of human video with no robot in the loop, finding a near-perfect scaling law where a single teleoperation demo sufficed to learn a new task after pretraining.
Reported metric: 20,000 hours
Reporting date: (not necessarily the event date)
Sources
Attribution and reporting linked by this record; links do not imply independent verification.
- Welcome to February 26, 2026 (theinnermostloop.substack.com)
Cite or correct this record
Suggest a correction — opens a GitHub draft for you to review and submit.