RoboDojo leaves robot models at 12.8%

RoboDojo’s unified benchmark spans 42 simulation tasks and 18 real-world tasks, yet its top AI model reached only 8.80% and 12.8% success respectively, versus 76.03% and 100% for human experts, HKU says. The benchmark paper combines generalisation, memory, precision, long-horizon execution, and open-vocabulary instruction following instead of letting a polished simulation score stand in for deployment reliability. Robot teams should add a sim-and-real gate to their evaluation plan, building on the sim-to-real robotics work already in the archive, before treating a demo as a product milestone.