SenseTime, together with teams from multiple universities and companies, arranged a major exam for OpenAI's GPT-6-Astra focused on "embodied intelligence." The exam was broad in scope—covering 12 simulation benchmarks and 8 real robot platforms—aimed at seeing clearly whether this star model can truly grow hands and feet from being a "brain" on a screen and get actual work done. The conclusion is a bit awkward: when given some special accommodations (using in-context learning and relaxing action step limits), GPT-6-Astra posted near-perfect scores on many traditional embodied benchmarks; but once it was placed on real robots, in scenarios demanding genuine hands-on operation such as fine contact tasks and long-horizon tasks, its error rate climbed.
The real problem is not "whether it can move," but "whether it knows whether it moved correctly." The research points out that GPT-6-Astra often misjudges success states in physical manipulation—it thinks it has grasped or placed an object properly when in fact it may not have; and once it goes wrong, remediation is quite difficult, plus the trial-and-error cost of real-robot execution is frighteningly high, leaving it a considerable distance away from being truly deployable. The team's overall judgment is: GPT-6-Astra has crossed the threshold of embodied capability, but there is still a clear gap before practical deployment.
This evaluation also incidentally punctured the bubble within the existing embodied evaluation system itself. The SenseTime team believes traditional benchmark scores are inflated because they are too far from real household needs, and evaluation should be adjusted toward directions closer to everyday life. In response, the team offered its own solution—the SenseNova-RoboRSI embodied system framework. Its cleverness lies in not needing to update base model weights; instead, it relies on execution feedback to let robots undergo "recursive self-evolution," continuously optimizing performance; this approach has already been validated in multiple simulation benchmarks. In one sentence: when large models score full marks on exams yet fumble in reality, what the industry may need to fix is not just the models, but also the ruler used to measure them.