Sensetime embodied-AI team published a study that reruns GPT-6-Astra across 12 simulated benchmarks and 8 robot bodies, covering table-top arms, humanoids, wheeled arms and dexterous hands. In RoboDojo, nine tasks that failed all five attempts under the standard protocol passed all five once two constraints were lifted: the model could study training-set demonstrations in context, and the step cap was raised to as much as 10x each benchmark official limit. Under those conditions RoboDojo success reached 81.9 per cent, and the LIBERO and SIMPLE suites saturated. The report also prices the runs: stacking bowls averaged about 131 minutes and roughly 19 dollars of inference, and putting a pen into a cup about 263 minutes and 55 dollars, at an assumed 1M tokens per dollar. Weak spots remain on BEHAVIOR large-scene wheeled-arm work and LHPR-VLN long-horizon navigation, plus failures to confirm whether a task is finished or to track physical state. Alongside the report, Sensetime open-sourced the SenseNova-RoboRSI harness and a 27B embodied model, claiming task-completion cost at 1/50 of the GPT-6-Astra figure, and describes a System 1 and System 2 pairing in which checking execution results, not only planning, drives performance.
Our takeThe methodological finding outlasts the scores: much of what leaderboards read as cannot is really ran out of steps. Lift the cap, allow in-context learning from demonstrations, and several suites saturate, meaning those benchmarks stop discriminating. What replaces them is cost per completed task plus closed-loop verification: four hours and 55 dollars for a pen-in-cup run is a compute-budget statement, not a capability one. The 1/50 cost claim and the weak cross-task transfer are vendor-reported, but harness and weights are public, so independent reruns are the real test.