SenseTime published a study that re-evaluates GPT-6-Astra on embodied robot tasks across 12 simulated benchmarks, covering tabletop arms, humanoid robots, wheeled-arm platforms and dexterous hands, with eight robot types in the mix. The headline finding is about test settings rather than weights: on nine RoboDojo tasks that failed five times out of five, the model reached five-out-of-five success after the company lifted two constraints without any retraining. The first constraint was access to training demonstrations, allowing in-context learning; the second was a step limit raised to as much as ten times each benchmark's official setting. Under those conditions GPT-6-Astra scored 81.9% on RoboDojo, and SenseTime says its results on LIBERO and SIMPLE are saturated. Cost is reported separately: stacking bowls averaged about 131 minutes and roughly $19 of inference, while putting a pen into a holder averaged about 263 minutes and $55. SenseTime says the weakest areas remain BEHAVIOR, a large-scene wheeled-arm suite, and LHPR-VLN, a long-horizon visual navigation benchmark. It lists failure modes that persist even at higher budgets, including an inability to confirm whether a task is finished, poor handling of physical state changes, and losing track among subtasks. The company also examined pairing a slow System 2 model (GPT-6-Astra) with a fast System 1 policy, and reports that delegating to the fast model without checking execution results cuts cost but damages performance. It released a framework called SenseNova-RoboRSI, a 27B embodied model and a technical report, claiming task cost at one-fiftieth of GPT-6-Astra's.
Our takeThe step-limit experiment is the part worth keeping: some of what reads as a capability failure is a search-budget failure, which means leaderboard numbers depend on a knob most papers do not report. The cost figures make the same point from the other side, at 263 minutes and $55 for one pen-in-holder trial. Until harness settings, verification loops and dollar cost are published next to scores, an embodied benchmark measures a configuration rather than a model.