Using Neural Networks to Predict the Normal Reactions of a Walking Robot

Purpose of reseach. This work is devoted to solving one of the problems associated with the control of walking robots based on their dynamic mathematical model − the presence in it of obvious mechanical bonds due to reactions of bonds with the supporting surface. To solve this problem, it is proposed to use a fully connected neural network to evaluate the forces of normal reactions between the surface and the feet of a bipedal walking machine during its implementation of one step.

Methods. The paper considers two neural network architectures based on fully connected layers with ReLU activation functions. The architecture of the neural network includes five fully connected layers (input, output and three hidden), and an alternative architecture includes a thinning layer after each fully connected layer. The input data for the network are the state of the robot and the required control actions, and the output is the predicted reaction forces. The training sample is generated by modeling a complete dynamic model of the robot. The network is built and trained using machine learning libraries Keras and TensorFlow.

Results.The generation of training sample for neural network is described here, and it is carried out the training of two architectures of neural networks. Based on the simulation data, it was established that both trained neural networks are able to accurately predict the values of normal reactions using the values of generalized coordinates and velocities, as well as control actions as input, however, a static prediction error is observed.

Conclusion. The results obtained within the framework of the article can be further used to control the movement of bipedal walking machines on various types of surfaces.

1. Werner A., Henze B., Rodriguez D.A., Gabaret J., Porges O., Roa M.A. Multi-contact planning and control for a torque-controlled humanoid robot. Intelligent Robots and Systems (IROS): Proc. IEEE/RSJ Intern. Conf., Daejeon, South Korea, 2016,pp. 5708-5715.

2. Posa M., Cantu C., Tedrake R. A direct method for trajectory optimization of rigid bodies through contact. Intern. J. of Robotics Research, 2014,vol. 33(1),pp. 69-81.

3. Jatsun S., Savin S., Yatsun A. Parameter optimization for exoskeleton control system using sobol sequences. Symposium on Robot Design, Dynamics and Control. Springer, Cham, 2016,pp. 361-368.

4. Featherstone R. Rigid Body Dynamics Algorithms. Boston, MA: Springer US, 2014, 271 p.

5. Mason S., Righetti L., Schaal S. Full dynamics LQR control of a humanoid robot: An experimental study on balancing and squatting. Humanoid Robots: Proc. IEEE-RAS Intern. Conf., Madrid, Spain, 2014,pp. 374-379.

6. Savin S., Jatsun S., Vorochaeva L. Modification of constrained LQR for control of walking in-pipe robots. Dynamics of Systems, Mechanisms and Machines (Dynamics): Proc. IEEE Intern. Conf., Omsk, Russia. 2017,pp. 1-6.

7. Savin S., Jatsun S., Vorochaeva L. State observer design for a walking in-pipe robot. MATEC Web of Conferences: EDP Sciences. 2018,vol. 161,pp. 03012.

8. Schmidhuber J. Deep learning in neural networks: An overview. Neural Networks, 2015,vol. (61),pp. 85–117.

9. Srivastava N., Hinton G., Krizhevsky A., Sutskever I., Salakhutdinov R. Dropout: a simple way to prevent neural networks from overfitting. J. of Machine Learning Research, 2014,vol. 15(1),pp. 1929-1958.

10. Abadi M., Barham P., Chen J., Chen Z., Davis A., Dean J., … Kudlur M. Tensorflow: A system for large-scale machine learning. Operating Systems Design and Implementation: Proc. 12th Symposium, Savannah, GA, USA, 2016,pp. 265-283.

11. Glorot X., Bengio Y. Understanding the difficulty of training deep feedforward neural networks. Artificial Intelligence and Statistics: Proc. of the 13-th Intern. Conf., Scottsdale, AZ, USA, 2010,pp. 249-256.