HumanoidBench: Simulated Humanoid Benchmark for Whole-Body Locomotion and Manipulation
HumanoidBench is a simulated benchmark of 27 whole-body tasks for a humanoid robot with dexterous hands: 12 locomotion tasks and 15 whole-body manipulation tasks. It runs in MuJoCo, with a Unitree H1 fitted with two Shadow Dexterous Hands as the primary robot. The authors report that off-the-shelf reinforcement learning methods fall below the success threshold on most tasks.
Whether a learning method can control a high-dimensional humanoid to walk, run, stand, balance and climb stairs, and to carry out whole-body manipulation such as pushing, opening doors and cabinets, and moving packages. Performance is reported as episode return per task.
Each task is a MuJoCo environment with dense and sparse reward terms, and the return is the sum of rewards over an episode. The paper compares DreamerV3, TD-MPC2, SAC and PPO (PPO only on a subset of tasks) and reports learning curves averaged over three seeds rather than a single normalized score; success is shown qualitatively against a per-task threshold. Some tasks have simple and hard variants. A hierarchical method that reuses pre-trained walking and reaching policies is evaluated on push and package.
Testing humanoid whole-body control on hardware is costly and fragile, so a shared simulated suite lets researchers compare methods before transfer to real robots. The reported results show flat reinforcement learning struggling on many whole-body tasks, with hierarchical use of low-level skills performing better.
Scoring
| Metric | Key | Unit | Direction |
|---|---|---|---|
| Episode return (sum of per-step rewards) | return | Unitless | Higher is better ↑ |
Hardware
Primary robot: simulated Unitree H1 with two Shadow Dexterous Hands on the arms. The repository also supports the Unitree G1 with three-finger hands and a low-dimensional-hand H1 variant. The paper also lists the Agility Digit as a supported body; it is not linked here.
Recorded outcomes
No result files yet
Evidence
Related evaluations