Neural Network-Based Joint Angle Prediction for the NAO Robot

One common problem in closed-loop robot motion control is the motor response delay. When computing, for instance, a walking step for a humanoid robot, the available state of the sensor joints comes from the past and the newly computed actuator positions will not be set immediately. This might lead to suboptimal motion planning, especially for fast and dynamic motions. In this paper, we present an approach for bridging this time gap and predicting the state of the joints of a NAO robot by using neural networks. The training of the neural networks was based on a dataset that covers multiple full tournaments played by the B-Human SPL team.