UBTECH Tests GPT-6 Astra in Six Tasks; Direct Grasping Falls Short

UBTECH recently disclosed six real-robot and simulation tests of its GPT-6 Astra robot. The real-world trials covered grasping a Rubik's cube, grasping a lucky bag and autonomous obstacle avoidance. The simulation tests covered basket transport, direct-control bowl stacking and world-model-assisted bowl stacking.

The six tests were designed to compare three technical paths: direct control by Astra; calling UBTECH's in-house Thinker-VLA model to execute operations; and combining Thinker-WM to predict action outcomes and assist decision-making. UBTECH said the setup spans perception, decision-making and execution, and reflects its effort to benchmark foundation models on both real hardware and simulated environments.

According to the disclosure, direct grasping was not successful, while the robot completed the bowl-stacking task with the collaborative model. UBTECH says the comparison highlights capability differences between direct control and model-assisted execution, offering a reference for improving how VLA and world models work together in embodied AI systems. The results suggest direct end-to-end control remains limited for certain manipulation tasks, while model collaboration can improve task completion.