PAKE: Learning Whole-Body Loco-Manipulation with Partial Kinematic Embeddings
This work proposes a framework for whole-body loco-manipulation by splitting it into two parts: generating partial kinematic references and low-level imitation control. It uses a Kinematic Normalizing Flow (KNF) to create an embedding of kinematic redundancy, allowing a high-level controller to select the best motion. A low-level policy then ensures these motions are dynamically feasible, leading to high-precision tracking.