A robot model trained on 30,000 hours more than doubled its hit rate.
Chinese firm AgiBot published a paper on 4 September testing whether robot models improve with data like chatbots do. Its model finished 17.1% of unseen tasks after 300 hours of training and 44.1% after 30,000. Nineteen of twenty skill groups improved, with no sign the curve had flattened.
Why it mattersRobot progress has been stuck on hand-built demos rather than anything that scales. If recorded hours reliably buy skill, home robots become a data problem, not a science one.
▶ Related video: Robot Data Scaling: Power Laws
Read in the app — free, in 9 languages
Related stories
A quarter of DeepMind's agent swarm blew the whistle on cheating.2026-09-15Ant threw its robot contest open to the world, hardware included.
2026-09-14A Turing winner says AI agents learn to lie because it pays.
2026-09-14Former EPA staff counted 30 rollbacks around AI data centres.
2026-09-14Giving a robot model a sense of touch lifted its success to 72.5%.
2026-09-13