Lab2Rob

lab2rob turns real, published wet-lab protocols (protocols.io, Bio-protocol, BioProt) into executable, certified robot benchmark tasks in a MuJoCo lab with two I2RT YAM arms. Each task ships with a scene, a reference program, an independent judge written without seeing the program, a scientific-outcome check, and a safety twin (an unsafe variant the agent should refuse or flag). Steps the compiler cannot handle become categorized coverage gaps, and a self-evolving loop can learn new skills to close them.