SJTU AutoLab's A²-MEAS Hits 2.21x BEHAVIOR 2026 Score, 93.98% on RoboTwin With Code-as-Correction Agents

Shanghai Jiao Tong's AutoLab released A²-MEAS, an agentic embodied-intelligence framework that uses generated correction code as the interface between a MetaBrain planner, diagnostic agents, and a 2B-param Qwen3-VL action model. It scores 0.5755 on the BEHAVIOR 2026 50-task public subset (2.21x the 2025 winner), 0.6411 on the 100-task full set, 93.98% average on RoboTwin 2.0, and 57.92% on real-robot tasks vs 26.67% for baseline. The group is open-sourcing weights, training scripts and infrastructure.