Speeding Up Inverse Kinematics Through Learning

Abstract

A fast and efficient inverse kinematics solution is central to robotic planning. Tradi- tionally, algorithms for solving the inverse kinematics were based on root-search and optimisation algorithms. This approach can be accelerated by using neural networks to estimate an initial guess. That starting point can then be refined with only a few iterations of a numerical solver. Machine learning is already being used for this purpose. However, the existing approaches can only produce collision-free solutions to the inverse kinematics in the environment in which they were trained. Supervised training is often used to learn the inverse kinematics, but this requires generating a large amount of training data in advance. We introduce an unsupervised training method that removes the need for this initial data generation. Moreover, our approach also works in envi- ronments that were not seen during training. We analyse the problems that can arise when learning the inverse kinematics with neural networks in detail. To address these issues, we present a twin-headed network architecture, a singularity-free joint angle representation, and a boosting procedure that improves performance on inputs that are rarely seen during training. We validate our approach in complex, three-dimensional environments using the humanoid robot Agile Justin. Training the network in this setting takes just 6.9h with our approach, compared to 95.4h of data generation plus 5.4h of training when using supervised learning. Our model achieves an average position error of 10.8cm, compared to Agile Justin’s reach of 1.85m, and an average orientation error of 0.087 radians. On top of this, it avoids obstacle collisions completely for 99.1% of target poses. By using the network’s prediction as an initial guess for a numerical solver, a collision-free solution to the inverse kinematics can be computed quickly and accurately.