Zhihui's 6 humanoids staff a real convenience store for 24h on livestream: 4 pass probation

机器之心 · wechat · 2026-11-03

Zhihui Robotics ran a 24-hour livestreamed stress test in a real Shanghai convenience store: six Elf G2 humanoids divided the work (greeting, restocking, shelf patrol, coffee, checkout, cleaning) and were graded like real employees on hours, throughput, customer reviews and task success. Four passed probation; the cleaner and shelf-patrol units failed on engagement and stability.

Three real-store challenges: soft-package grasping requiring real-time force/tactile adjustment; long-horizon restocking with interdependent subtasks and state memory; obstacle avoidance under strict safety constraints with unpredictable customers — plus testers deliberately setting traps (blocked paths, shaken boxes, sugared vs. diet cola).

GenieMaster embodied Agent framework: a shared "brain" closing the judge-act-feedback loop, attributing failures and adjusting strategy autonomously. In earlier tests, food-packing/picking/restocking success rose from 28.6% on day one to 85.7% by day five; trajectories feed daily summaries into long-term memory and training data, reducing reliance on human teleoperation.

The piece also recaps Zhihui's seven releases over ten days (AGIBOT WORLD dataset v4, GenieStudio 2.0, G2MAX, GenieMaster, GenieSim 3.5, GO-3 Preview, collective experience learning) and argues embodied AI deployment requires full-stack coordination.

Related event: AgiBot's Six Humanoid Robots Run Shanghai Convenience Store for 24 Hours(2 posts)→

Original post →

More from Embodied

Embodied channel →