Whole-body loco-manipulation—coordinating a robot's legs and arm simultaneously—is difficult because of high-dimensional action spaces and redundant degrees of freedom. The PAKE framework, described in a paper on arXiv, tackles this by splitting the problem into two parts: generating partial reference motions and then imitating them with a low-level controller. A Kinematic Normalizing Flow (KNF) model produces diverse partial kinematic references, and a trained high-level controller selects among them in latent space to exploit the robot's redundancy.
In simulation, PAKE outperformed the evaluated baselines on a 4,000-target benchmark in both tracking accuracy and feasible workspace coverage. For hardware, the authors tested the approach on a quadrupedal robot equipped with a six-degree-of-freedom robotic arm across 24 episodes spanning 8 mobile loco-manipulation tasks. They report end-effector pose-tracking errors of 4.5 cm and 0.14 rad, with linear and angular velocity errors of 0.1 m/s and 0.01 rad/s for locomotion tracking.
The authors describe the method as a practical solution for accurate and generalized whole-body control in high-DoF robotic systems. As a single-source digest, there are no contrasting findings to compare; the claims are reported as presented in the paper.