SkillBlender performs versatile autonomous humanoid loco-manipulation tasks within different embodiments and environments, given only one or two intuitive reward terms.
SkillBlender performs versatile autonomous humanoid loco-manipulation tasks within different embodiments and environments, given only one or two intuitive reward terms.
Overview of SkillBlender. We first pretrain goal-conditioned primitive expert skills that are task-agnostic, reusable, and physically interpretable, and then reuse and blend these skills to achieve complex whole-body loco-manipulation tasks given only one or two task-specific reward terms.
Our SkillBench is a parallel, cross-embodiment, and diverse simulated benchmark containing three embodiments, four primitive skills, and eight loco-manipulation tasks.
Qualitative comparison between different methods. Our SkillBlender not only achieves higher task accuracy, but also avoids reward hacking and yields more natural and feasible movements.
Visualization of whole-body per-joint weights at different task stages. More blue means more Reaching, and more green means more Walking. This visualization highlights the spatial-temporal decomposition of our skill blending, where the two skills interleave rather than one skill dominating the overall motion.
We also provide a sim2real toolkit support that deploys our simulation-trained primitive skill policies in the real world using a Unitree H1 humanoid robot.
