AGIBOT Unveils GO-3 Preview Robot Model: Predicts Outcomes Before Acting, Rankings Self-Reported
Reported by Tech Times1 source ↓
AGIBOT unveiled GO-3 Preview on Oct. 29 in Shanghai. The embodied AI model forecasts how a scene may change before selecting robot actions; its claimed benchmark rankings have not been independently confirmed.
Key points
- AGIBOT says GO-3 Preview ranks first on RoboDojo, RoboTwin and RoboColiseum
- The model is not open-source; AGIBOT says it plans to release it by the end of 2026
- AGIBOT reports a phone-packaging task's success rate rose from 50.8% to 98.3% with fleet training
Details
GO-3 Preview is designed for AGIBOT robots and is based on what the company calls a World Reasoning Action Model. It predicts future visual states from a camera view and task, then uses the forecast and the robot's state to generate actions. AGIBOT describes the approach as adding a prediction step between perception and action.
AGIBOT's Nov. 3 release claimed first place on RoboDojo, RoboTwin and RoboColiseum. Tech Times reports those rankings were supplied by AGIBOT; it found no visible entries on the RoboDojo leaderboard page it checked, and says outsiders cannot reproduce the results without public model weights and an evaluation protocol. PingWest reported that AGIBOT had not published a technical report.
AGIBOT says its post-training system uses a fleet of 100 physical robots alongside simulation and world models, pooling successful and failed attempts.
Written from coverage as of
Images
- Tech Times