SenseNova-RoboRSI reports higher robot agent scores by evolving harnesses, weights unchanged

SenseNova, the SenseTime team, reports that SenseNova-RoboRSI raises the RoboDojo average score of an embodied AI agent to 56.83, against a published GPT-6 Astra baseline of 28.97, while leaving the model weights unchanged.

The system reaches a 50.83% task success rate using a recursive self-improvement loop that evolves the agent harness from physical task feedback, plus multi-point end-effector prediction and Planning and Feedback Subagents. The figures come from the team's announcement on X, relayed by Rohan Paul.

The team says the code will be open-sourced at github.com/OpenSenseNova/SenseNova-RoboRSI, with no release date given, and that a detailed technical report is forthcoming. Until both appear, the results cannot be checked against the methods.