HKU Researchers Develop Unified Benchmark for Physical AI
aiinsider ·
Researchers led by the University of Hong Kong have developed RoboDojo, a unified benchmark for evaluating physical AI systems across both simulated and real-world robot tasks. The platform includes 30 robot policies, 42 simulation tasks, and 18 real-world tasks, and it tests capabilities such as generalization, memory, precision, and multi-step execution. It was created with collaborators from nearly 20 universities, including UC Berkeley and Tsinghua University. Early results show a large performance gap: the top AI model achieved success rates of 8.8% in simulation and 12.8% in real-world tests, while human experts scored 76.03% and 100%, respectively. The benchmark aims to provide a standardized, reproducible way to compare robot-learning methods across environments, rather than relying on isolated demonstrations. RoboDojo's open-source resources have been downloaded more than 100,000 times from Hugging Face and received over 100,000 views on X in its first week.
aiinsider