LOInK: Learned Optimal Inverse Kinematics via Structured Neural Surrogate Models

This paper presents LOInK (Learned Optimal Inverse Kinematics), a machine learning approach for generating approximately optimal solutions to inverse kinematics problems in robotics. The method trains on sampled configurations with associated task variables and a specified cost function. LOInK learns a bi-Lipschitz invertible mapping between configuration space and a decoupled task/latent space. A key innovation is that the latent space is structured so that cost-minimizing solutions are positioned at the origin, enabling efficient sampling of optimal configurations.