UBTECH GPT-6 Astra Robot Fails Direct Control Bowl Stacking Test, Succeeds with Thinker-VLA and Thinker-WM Collaboration

On April 18, 2026, UBTECH disclosed six mixed real-machine and simulation tests for its GPT-6 Astra robot. In a direct control task to stack three ceramic bowls, the robot failed due to a 0.3-second latency in posture adjustment, causing the second bowl to slip and the third to fall off the table. However, when switched to Thinker-VLA collaborative mode, which integrates UBTECH's self-developed vision-language-action model, the robot successfully stacked the bowls with an error less than 0.8 millimeters. Further tests showed that integrating Thinker-WM world model enabled predictive compensation for physical dynamics, reducing action planning time to 1.1 seconds. The tests highlight the limitations of large language models in physical interaction and demonstrate the value of collaborative systems incorporating small models that understand physical constraints.