Sebastian Otte, Adrian Zwiener, Richard Hanten and Andreas Zell
This project investigates inverse recurrent forward models for many-joint robot arm control. First, Recurrent Neural Networks (RNNs) are trained to predict arm poses. Due their recurrence the RNNs naturally match the repetitive character of computing kinematic forward chains. We demonstrate that the trained RNNs are well suited to gain inverse kinematics robustly and precisely using Back-Propagation Trough Time even for complex robot arms with up to 40 universal joints with 120 articulated degrees of freedom and under difficult conditions. The concept is additionally proven on a real robot arm. The presented results are promising and reveal a novel perspective to neural robotic control.

The following videos show movement trajectories of different robot arms converging towards random poses. The blue coordinate system represents the target pose, whereas the red one represents the reference frame of the end-effector tip. The trajectories directly reflect the iteration loop - following the gradient from the recurrent forward model. Note that for producing the videos the maximum step length (in configuration space) was reduced in order to yield a slower motion trajectories at 25 fps.

The video below demonstrates that our approach also works for real robot arms. The robot arm performs an 'eight' trajectory. For all trajectory points (in Euclidean space) the inverse kinematics were computed successively via the inverted recurrent neural network.