96 Humanoid Robots Learn From Each Other, Lifting Success Rates From 81% to 97.3%
CyberRobooo · x · 2026-10-25
Teams from Tsinghua University, AgiBot, and Infinigence AI tested their Collective Experience Learning (CEL) framework on up to 96 dual-arm AgiBot G2 robots across five real-world tasks.
How it works: robots share successful task recordings. When one struggles, the system searches other robots' successful runs for similar situations to improve the policy, then pushes updates fleet-wide. Humans provide initial demonstrations, but ongoing learning no longer needs expert corrections.
Key numbers:
- Average success across five tasks: 81% → 97.3%
- Phone packaging: 79.2% → 98.3%
- Scaling fleet from 16 to 96 robots: 86.7% → 98.3%
Commercially relevant: AgiBot shipped 8,600+ humanoid robots in H1 2026, and shared-experience learning could make growing fleets more reliable while lowering per-task costs.
More from Embodied
- 80-year-old credits Tesla FSD for driving 2,800 of 3,000-mile road trip — elonmusk · 2026-10-25
- Chaos Audio's Stratus AI pedal turns text prompts into custom guitar tones — BenBajarin · 2026-10-25
- Agility Robotics goes public Nov 2 via SPAC, ticker CCXI to become AGLT — RyanEvansAFT · 2026-10-25
- TechCrunch rounds up the hottest AI wearables: Muse Charm, Bee, Friend, Plaud — TechCrunch AI · 2026-10-25
- AgiBot's 110-robot sync demo proves fleet scheduling, not fleet readiness — VectorWireAI · 2026-10-25
- NEURA Robotics unveils 4NE1 Gen4 humanoid: 20% lighter, powered by in-house Neuron AI — CyberRobooo · 2026-10-25