SkillBench, introduced with SkillBlender: Towards Versatile Humanoid Whole-Body Loco-Manipulation via Skill Blending
SkillBench is a simulated, cross-embodiment benchmark of eight humanoid loco-manipulation tasks, introduced with the SkillBlender framework. It covers three Unitree humanoids (H1, G1 and H1-2) and four primitive skills. It tests whether pre-trained, goal-conditioned skills can be blended into complex whole-body tasks with little task-specific reward engineering.
Task accuracy and motion feasibility for whole-body loco-manipulation, such as reaching, pressing buttons, pushing boxes, carrying packages and shooting a ball, across humanoid bodies. The paper also argues that its method avoids reward hacking.
The eight tasks are FarReach, ButtonPress, CabinetClose, FootballShoot, BoxPush, PackageLift, BoxTransfer and PackageCarry. Policies are built by pre-training task-agnostic primitive skills and blending them with a hierarchical reinforcement learning framework. The paper states a set of evaluation metrics that balance accuracy and feasibility; their exact definitions were not verified here.
Whole-body loco-manipulation, meaning walking while using the hands, is a core humanoid capability, and SkillBench targets three humanoid bodies directly. It is a recent, single-paper benchmark, so its adoption and long-term maintenance are not established.
Scoring
| Metric | Key | Unit | Direction |
|---|---|---|---|
| Task accuracy | task_accuracy | Unitless | Higher is better ↑ |
Hardware
Three Unitree humanoids: H1, G1 and H1-2, mapped to the roster slugs unitree-h1, unitree-g1 and unitree-h1-2. Degrees of freedom for each embodiment are not given on the project page or in the abstract.
Recorded outcomes
No result files yet
Evidence
Related evaluations