A partially recurrent neural network to perform trajectory planning, inverse kinematics, and inverse dynamics

This paper proposes a three-layer partially recurrent neural network to perform trajectory planning, solve the inverse kinematics and the inverse dynamics problems in a single processing stage. The feedforward structure of the neural model entails fully connected layers. The feedback links consists in output-input and input-input connections. All the connections are trainable by error backpropagation with variable learning rate and momentum. The network generated trajectories for the PUMA 560 manipulator. The tests comprise generation of four different types of trajectories. Each path is provided in spatial positions, joint angles and joint torques. The results suggest that the model is able to yield trained trajectories given only their initial and final points. Moreover, the results suggest that the model is robust to noise in the trajectories with lower level of complexity.