SenseNova-RoboRSI: doubles RoboDojo score without changing model weights
SenseNova open-sourced RoboRSI, a self-improving agent harness that raises the same model's RoboDojo score from 28.97 to 56.83 without changing base weights.
Same model, nearly double the RoboDojo score. SenseNova-RoboRSI is exploring a new path for improving Embodied AI. Instead of training a larger model, it focuses on making existing models more effective through a self-improving agent harness. ▪️ RoboDojo: 56.83 average score, 50.83% success rate ▪️ LIBERO-PRO: 94.50 ▪️ RoboCasa: 64.60 On RoboDojo, the score jumps from the public GPT-6 Astra baseline of 28.97 to 56.83, a 96.2% relative improvement, without changing the base model's weights. At the heart of SenseNova-RoboRSI is recursive self-improvement at the harness level. It analyzes execution traces, identifies weaknesses, explores alternative strategies, and repeatedly evaluates the results before incorporating successful improvements into future versions of the agent. Rather than updating the model itself, SenseNova-RoboRSI optimizes how it plans, calls tools, executes actions, and learns from physical feedback. Excited to see where this next step in Embodied AI takes us! SenseNova-RoboRSI GitHub repository is live, with the full code and technical report coming soon: http://github.com/OpenSenseNova/SenseNova-RoboRSI
This post quotes another post: x.com/sensetime_ai/status/2108929659912221058