GPT-6 Astra directly takes over the robot "brain," with simulation scores surpassing dedicated VLA.

Beating AI News Flash: The RoboDojo team, a robotic manipulation benchmarking project led by the University of Hong Kong's MMLab, conducted an unconventional experiment. Instead of pairing GPT-6 Astra with a trained robotic policy, they let this general-purpose multimodal model directly view the camera, read the robot's state, understand the task, and decide on its own how to move next. In between, there was only a fixed set of control tools that converted the target positions given by Astra into robotic arm actions.

Across 42 simulated tasks, Astra scored 28.97 points with an average success rate of 22.48%, surpassing the previous public leader DM0.5's 24.90 points and 19.34%. When the same control setup was switched to GPT-5.5, it scored only 1.13 points and 0.88%.

The RoboDojo public leaderboard had previously been dominated mainly by specially trained robotic policies such as VLA and WAM. Now, some of the motion decision-making that originally required specialized models can be taken over directly by general-purpose large models, and the overall results even exceed the previous leader. For the first time, the question of whether robots still need a separately trained "brain" has become very concrete.

But Astra still cannot replace the robot's "cerebellum." It is noticeably weaker at tasks that depend on real physical control, such as insertion, pouring liquids, and fine contact. Real-robot testing was also halted early by the team after dangerous actions occurred multiple times and damaged hardware.

As general-purpose models become more capable, the division of labor in robotic models may be redrawn: large models handle understanding the environment, judgment, and planning, while specialized VLA and traditional controllers handle making movements accurate and stable.