SenseTime open-sources 27B robot model matching GPT-6-Astra at ~1/50 the cost

rohanpaul_ai · x · 2026-10-24

SenseTime released the most comprehensive empirical study of GPT-6-Astra's embodied intelligence, spanning 12 simulation benchmarks and 8 real-robot platforms, alongside the open-source SenseNova-RoboRSI harness and a 27B model, RoboRSI-27B-Preview.
Under a unified protocol (ICL + up to 10x interaction budget), GPT-6-Astra hits clear SOTA on all 12 benchmarks: RoboDojo jumps from 27.6% at 1x to 81.9% at 10x, DexJoCo reaches 87.3% vs. 40.3% prior SOTA, and LIBERO is pushed to saturation.
But it's expensive: even placing pens in a holder costs $55 in inference. The key idea is experience internalization — distilling embodied experience (visual examples, robot states, action trajectories, tools, recovery strategies) into the open model, which achieves active perception, spatial reasoning, planning and feedback-driven revision at roughly 1/50 the task cost.

More from Embodied