Generating Dual-Arm Inverse Kinematics Solutions using Latent Variable Models

Solving the Inverse Kinematics Problem is a fundamental skill for humanoid robots and their interaction with the world. In contrast to industrial manipulators, humanoid robots can simultaneously grasp or push objects with two hands. This demands generating self-collision-free inverse kinematics solutions for both arms in a minimal time. Recent research in the context of single-arm manipulators utilizes deep generative models to obtain a whole set of feasible solutions. This paper investigates their performance on the more complex dual-arm problem. To this end, we extend the problem space to a dualarm setup and learn inverse kinematics solutions, providing the two target end effector poses as a conditional variable. We propose an approach based on Conditional Variational Autoencoders and investigate the trade-off between model size, accuracy, and its benefit when being used for seeding a numeric solver. In this context, we also evaluate the influence of flexible learning-based priors against fixed Gaussian priors. Our approach can initialize the solver within 1 ms and drastically increase the number of returned solutions. The results show that even less accurate models can drastically increase the performance of a numeric solver while yielding significantly shorter solving times compared to a state-of-the-art flow-based method.