AGIBOT's GO-3 Preview tops three embodied AI benchmarks

AGIBOT positions GO-3 Preview as the next step after its earlier GO-1 and GO-2 models. GO-1 introduced the ViLLA framework, and GO-2 pioneered an action chain-of-thought approach. The new model extends scaling beyond pre-training to cover post-training and learning after deployment, an approach AGIBOT calls full-lifecycle scaling.

The system pairs the model with two supporting pieces: a hybrid post-training system that combines a fleet of 100 physical robots with physics simulation and world models, and an embodied agent framework that standardizes how tasks are planned, executed and reviewed. In demonstrations, AGIBOT says a robot practiced tabletop curling for a few hours and achieved a higher win rate than typical human players, and that a robot in a real convenience store identified task failures and adjusted strategies without human demonstrations or manual tuning. AGIBOT describes that convenience-store result as Physical RSI, or recursive self-improvement in the physical world.

The company's broader footprint provides context for the release. AGIBOT says its portfolio spans humanoid robots, quadrupeds, dexterous systems and commercial cleaning solutions, and that its 20,000th robot rolled off the production line in September 2026. The first-place benchmark rankings and the planned open-sourcing by the end of 2026 are likely to shape how the model is evaluated outside the company.