KineNN : Kinematic Neural Network for inverse model policy based on homogeneous transformation matrix and dual quaternion

The modeling and control of a robot manipulator can be challenging considering different robot architectures
and different tasks. In this paper, we introduce a novel framework for data based control of robot operating
tasks using a novel, invertible neural network called Kinematic Neural Network (KineNN). To this end, we present two KineNN architectures based on the Rigid Body Transformation in the form of either the Homogeneous Transformation Matrix (HTM) or Dual Quaternion (DQ). The KineNN serves two purposes in our approach. First, it acts as the forward kinematic model of a robot within a model based reinforcement learning architecture where the output is the end effector position and orientation of the robot manipulator with given joint angles of the robot. Second, KineNN’s inverted architecture is used within the policy network making the policy network aware of the actual robot architecture, which allows for an disentanglement of robot kinematics and task specific control resulting in improved training performance. Within the approach both policy and model NN share their parameters. The proposed framework was tested and evaluated on a Universal Robot (UR) 5. The results show that the architecture can successfully capture the robot kinematics and predict the world model state. The inverse model with shared parameters within the policy network outperforms a training without this sharing. We further conduct a transfer learning where we modify the arm lengths and number of joints. In this experiment, transferring KineNNs parameters yielded faster convergence in comparison to re-training a model from scratch.