On October 20, SenseTime open-sourced an embodied AI framework called SenseNova-RoboRSI. Its key selling point in one sentence: it cuts the cost of completing tasks to 1/50 of GPT-6-Astra’s.
First, what did it test? SenseTime ran systematic tests using GPT-6-Astra across 12 simulated benchmarks and 8 types of robots. The results were fascinating: all 5 attempts at 9 tasks in RoboDojo had originally failed. Without retraining, they simply let the system keep interacting and raised the step limit by 10 times—and all 5 attempts succeeded. Some things it “couldn’t do” may just be cases where we hit the stop button too soon.
But waiting indefinitely comes at a high cost: stacking bowls took an average of about 131 minutes and around $19 in inference fees; putting a pen into a pen holder took an average of 263 minutes and around $55. Clearly, robots like these are still a long way from everyday use.
The open-source 27B model SenseNova-RoboRSI-27B-Preview distills validated embodied AI experience into its parameters: a 20.83% success rate on RoboDojo tasks, a score above 90 on LIBERO-Pro, an improvement on RoboCasa365 from 26.0% to 46.5%, and a 73% reduction in token usage.
My take: the bottleneck for embodied AI has long ceased to be “can it do the task?” The real questions are “how much does one attempt cost, and how long does it take?” The value of SenseTime’s system isn’t its benchmark scores, but that it lets the community reproduce and compare results. Shifting from chasing scores to scrutinizing cost and reliability—that’s the right path for embodied AI to become practical.
#AIFrontier
$BTC
First, what did it test? SenseTime ran systematic tests using GPT-6-Astra across 12 simulated benchmarks and 8 types of robots. The results were fascinating: all 5 attempts at 9 tasks in RoboDojo had originally failed. Without retraining, they simply let the system keep interacting and raised the step limit by 10 times—and all 5 attempts succeeded. Some things it “couldn’t do” may just be cases where we hit the stop button too soon.
But waiting indefinitely comes at a high cost: stacking bowls took an average of about 131 minutes and around $19 in inference fees; putting a pen into a pen holder took an average of 263 minutes and around $55. Clearly, robots like these are still a long way from everyday use.
The open-source 27B model SenseNova-RoboRSI-27B-Preview distills validated embodied AI experience into its parameters: a 20.83% success rate on RoboDojo tasks, a score above 90 on LIBERO-Pro, an improvement on RoboCasa365 from 26.0% to 46.5%, and a 73% reduction in token usage.
My take: the bottleneck for embodied AI has long ceased to be “can it do the task?” The real questions are “how much does one attempt cost, and how long does it take?” The value of SenseTime’s system isn’t its benchmark scores, but that it lets the community reproduce and compare results. Shifting from chasing scores to scrutinizing cost and reliability—that’s the right path for embodied AI to become practical.
#AIFrontier
$BTC