AgiBot launches Genie Master agent for self-improving G2 robots, plans 24-hour store test
AgiBot released Genie Master, an embodied-agent framework that lets its G2 robots find the cause of failed tasks and adjust without human demonstrations while working, IT Home reported on Oct 27. AgiBot says G2 units running it have packed hot food, picked and restocked goods in a real convenience store, and that six G2 robots will run a 24-hour live store test on Nov 2.
AgiBot's own post was not available to us, so the claims are unconfirmed.
AgiBot released an embodied-agent framework called Genie Master on Oct 27, IT Home reported, citing the company. The framework is meant to let a deployed robot improve without people in the loop: according to AgiBot, a G2 robot running it in a real convenience store located the reasons for its own failures, changed strategy and stored the experience, improving on several fine-grained tasks without human demonstrations or tuning. AgiBot describes this as the first validation of what it calls physical recursive self-improvement outside simulation.
The tasks named are packing ready-to-eat food, fetching goods and restocking shelves. AgiBot also plans a 24-hour continuous live-streamed shift by six G2 robots in a convenience store on Nov 2. No success rates, error counts or customer names were given in the report, and we could not find the release on AgiBot's own news page, so the performance claims remain the company's as relayed by one outlet. The day before, AgiBot showed the G2 Max, a heavy-payload version for palletising and line feeding.
MEKAWIRE context: the Nov 2 stream follows AgiBot's earlier live production-line demonstration of the G2 on tablet assembly. A 24-hour unattended retail shift would be a harder public test than a factory line, because store items and layouts change through the day; the stream is the first chance to check the self-improvement claim against observable error rates.