FLAIR · project
GSC: Generative Skill Chaining: Long-Horizon Skill Planning with Diffusion Models
Long manipulation tasks require decisions that remain compatible across several skills. A locally successful action can leave the robot in a state from which the next skill cannot succeed. GSC addresses this dependence by planning over a sequence of skills jointly.
Composing skills into a plan
Each skill is represented by a diffusion model. At inference time, GSC combines these learned distributions and samples across the skill sequence in parallel, coordinating the transitions between skills while imposing geometric constraints. This allows skills learned separately to be reused in new task combinations.
What the experiments show
Simulation and real-robot experiments examine long-horizon manipulation, including dependencies between actions, constraint satisfaction, and replanning after perturbations. The contribution is a probabilistic way to turn a library of skills into coordinated plans, with demonstrations and implementation details available through the project and code links above.
Publication
Generative Skill Chaining: Long-Horizon Skill Planning with Diffusion Models
Utkarsh A. Mishra, Shangjie Xue, Yongxin Chen, and Danfei Xu
CoRL 2023