SenseTime has open-sourced SenseNova-RoboRSI, a framework for evaluating embodied AI systems, together with an embodied model called SenseNova-RoboRSI-27B-Preview. The release follows a study in which SenseTime tested the large model GPT-6-Astra on 12 simulated robot benchmarks and 8 physical robots. All results below come from SenseTime's own tests.
SenseTime says that when it gave GPT-6-Astra access to training demonstrations and raised step limits to as much as 10 times the official settings, the model saturated the LIBERO and SIMPLE benchmarks and reached an 81.9% success rate on RoboDojo. In one case, nine RoboDojo tasks that had failed in all five attempts succeeded in all five once the model was simply allowed to keep interacting.
The cost was high. In SenseTime's extended-budget tests, stacking bowls took about 131 minutes and roughly $19 in estimated inference costs, while putting a pen into a pen holder took 263 minutes and about $55. SenseTime also found that GPT-6-Astra remained weakest on long-horizon tasks, struggled to transfer skills across tasks, and had trouble with deformable objects and fine control on real robots.
The company then tested pairing a slow reasoning model with a fast action model. Its best combination reached 54% success across 42 RoboDojo tasks with a closed loop that checks each result, though SenseTime says cost remains an issue.
The open-source harness packages each benchmark and its dependencies into a separate Docker image and exposes environment interaction and native scoring through a unified HTTP API. SenseTime says researchers can run end-to-end evaluations from a configuration file, without setting up each simulator by hand.
The 27B model is meant to build tested experience into model weights so it needs less trial and error. SenseTime reports it scored 26.13 on RoboDojo with a 20.83% task success rate, topped 90 on LIBERO-Pro, and raised its RoboCasa365 success rate from 26.0% to 46.5% while cutting token use by 73%. The company says it completes embodied tasks at one-fiftieth of GPT-6-Astra's cost.
SenseTime also argues that future embodied benchmarks should be built around real human needs, require robots to learn by interacting, and count thinking time as world time, since a pan keeps heating while a model deliberates. None of the results have been independently verified.
SEE ALSO: SenseTime SenseNova U1.5 Brings 8B-MoT Native Unified Vision · Infinigence AI Open-Sources APXInf for Embodied Edge Inference