HumanoidBench

HumanoidBench is the first comprehensive benchmark designed specifically for whole-body humanoid robot control. Developed by researchers at the University of Edinburgh, it provides 30 standardized tasks in a unified MuJoCo simulation environment, equally split between 15 locomotion tasks and 15 manipulation/combined tasks. Using a simulated Unitree H1 humanoid with 28 degrees of freedom, tasks range from basic locomotion (walking forward, turning, stepping over obstacles) to complex manipulation (picking up objects, opening doors, placing items on shelves) and combined tasks that require simultaneous mobility and manipulation. The benchmark includes standardized reward functions, observation spaces, and baseline results using PPO, SAC, and imitation learning. HumanoidBench has become the standard evaluation platform for humanoid learning algorithms, enabling systematic comparison across methods for whole-body control.