From the source
SenseTime has open-sourced SenseNova-RoboRSI, a harness for reproducing embodied-AI evaluations, plus a 27B embodied model it says cuts task cost to 1/50 of GPT-6-Astra.
- The release follows a study of GPT-6-Astra across 12 simulator benchmarks and 8 real robots, which found that lifting step limits and allowing in-context learning pushed the model close to saturation on several leaderboards.
- In RoboDojo, nine tasks that had failed five times out of five succeeded five times out of five once the step cap was raised tenfold, without retraining.
- But the study puts the cost of that success in view: stacking bowls took about 131 minutes and roughly $19 of estimated inference, while putting a pen in a holder took 263 minutes and about $55.
- SenseTime says SenseNova-RoboRSI-27B-Preview scored 26.13 on RoboDojo at a 20.83% success rate, and on RoboCasa365 raised success from 26.0% to 46.5% while cutting task token cost by 73%.
- The harness packages benchmarks and their dependencies as Docker images behind a unified HTTP API, so researchers can reproduce end-to-end evaluations from a configuration file instead of deploying simulators by hand.