RoboDojo

RoboDojo is a benchmark for evaluating generalist robot manipulation policies across both simulation and real-world settings. It provides 42 simulation tasks covering 5 evaluation dimensions (generalization, memory, precision, long-horizon, open-vocabulary) and 18 real-world tasks. The benchmark supports heterogeneous parallel simulation in Isaac Sim and includes RoboDojo-RealEval for reproducible real-world evaluation. Integrates 30 policies through XPolicyLab with a public leaderboard at robodojo-benchmark.com. In August 2026, the University of Hong Kong (HKU) issued an official press release positioning RoboDojo as the first Hong Kong-led unified benchmark for embodied AI, setting new standards for robot reliability evaluation. The arXiv paper was updated to v3 on 2026-08-13. Xiaomi-Robotics-1 reports an average score of 20.07 on RoboDojo, significantly outperforming the prior state-of-the-art of 13.07.