Inherently Constraint-Aware Control of Many-Joint Robot Arms with Inverse Recurrent Models

In a recent study, it was demonstrated that Recurrent Neural Networks (RNNs) can be used to effectively control snake-like, manyjoint robot arms in a particular way: The inverse kinematics for control are generated using back-propagation through time (BPTT) on recurrent forward models that learned to predict the end-effector pose of a robot arm, whereby each joint is associated with a certain computation time step of the RNN. This paper further investigates this approach in terms of constraint-aware control. Our contribution is twofold: First, we show that an RNN can be trained to also predict the poses of intermediate joints within such an arm, and that these can consequently be included in the control-optimization objective as well, giving full control over the entire arm. Second, we show that particular components of the arm's target can be selectively switched on and off by means of "don't care" signals. This enables us to handle constraints inherently and on-the-fly, without the need of any outer constraint mechanisms, such as additional penalty terms. The experiments demonstrating the effectiveness of our methodology are carried out on a simulated three dimensional 40-joint robot arm with 80 articulated degrees offreedom.