SenseNova-RoboRSI-27B-Preview Robot Model Ties Into Open Simulator
The interesting part is the closed feedback loop: execution results feed back into the model's next decision, which makes failure analysis and model comparison reproducible instead of guesswork. Robotics engineers and researchers benchmarking vision-language-action models can reuse the loop with their own tasks. Against a named alternative, LIBERO-Pro's 96.1% success under direct control is the figure to test competitors against.
SenseTime's open project connects the SenseNova-RoboRSI-27B-Preview model to a simulated robot, where the model receives camera images, picks a movement, observes the simulator's feedback, and iterates. The model scores 26.13 average and 20.83% task success on RoboDojo. On LIBERO-Pro with direct control it reaches a 96.1% success rate. The setup lets developers compare models and diagnose why a robot fails a task.
You can use this open project to connect a model to a simulated robot.
Then, you can measure how well the robot completes a task.
The basic loop is simple:
1. The model, SenseNova-RoboRSI-27B-Preview, receives a task and images from the robot's cameras.
2. It chooses a movement
3. The simulator executes that movement and returns new observations
4. The model uses that feedback to decide what to do next
Some of the model highlights:
• Inspects visual details and reasons about spatial relationships • Adjusts actions based on execution feedback • 26.13 average score and 20.83% task success on RoboDojo • 96.1% success rate with direct control on LIBERO-Pro
This is pretty useful for comparing models and understanding why a robot failed.
