Joint Angle Error Reduction for Humanoid Robots Using Dynamics Learning Tree
In this paper, we discuss two problems with joints of low-cost humanoid robots.
The first problem is communication errors occurring in angle sensors. We propose a method of compensating for the sensor values by using estimated sensor values by learning the corresponding relationships between the command and sensor values.
Second, there are errors between the command and sensor values. The degree of such errors in a robot arm is affected by both gravity and joint-motion directions. By learning the corresponding relationships between these two factors and the errors, we can estimate these errors and use this estimation to reduce motion error. One of the distinguishing points of the proposed methods is that these two problems are solved by adaptive learning that works under the background system of a moving robot. Another distinguishing point is that the proposed method adapts to the specifications of a robot’s joints regardless of intensive a priori knowledge about the specifications.
From experimental results, we found that it is possible to infer the necessary value to compensate the sensor values that occur in the event of communication error. Moreover, by estimating the error between the command and sensor values and using this estimation to reduce the error, we succeeded in reducing the error in joint angle.
- 1.Harmonic Drive. https://www.hds.co.jp/products/lineup/hd/
- 3.Roumeliotis, S.I., Sukhatme, G., Bekey, G.A.: Fault detection and identification in a mobile robot using multiple-model estimation. In: Proceedings IEEE International Conference Robotics and Automation, Lueven, Belgium, pp. 2223–2228 (1998)Google Scholar
- 7.He, C., Zhang, Y., Meng, M.: Backlash compensation by neural network online learning. In: Proceedings IEEE International Symposium on Computational Intelligence in Robotics and Automation, pp. 161–165 (2001)Google Scholar
- 8.Baruch, I.S., Beltran, R.L., Nenkova, B.: A mechanical system backlash compensation by means of a recurrent neural multi-model. In: Second IEEE International Conference on Intelligent Systems, pp. 514–519 (2004)Google Scholar
- 9.Shibata, T., Fukuda, T., Tanie, K.: Nonlinear backlash compensation using recurrent neural network. Unsupervised learning by genetic algorithm. In: Proceedings of 1993 International Conference on Neural Networks, IJCNN-93-Nagoya, Japan, vol. 1, pp. 742–745 (1993)Google Scholar
- 10.Numakura, A., Kato, S., Sato, K., Tomisawa, T., Miyoshi, T., Akashi, T., Kim, C.H.: FAD learning: separate learning for three accelerations-learning for dynamics of boat through motor babbling. In: Proceedings of IEEE International Conference on Robotics and Automation, pp. 5609–5614 (2016)Google Scholar
- 11.Watanabe, K., Nishide, S., Gouko, M., Kim, C.H.: Fully automated learning for position and contact force of manipulated object with wired flexible finger joints. In: Fujita, H., Ali, M., Selamat, A., Sasaki, J., Kurematsu, M. (eds.) IEA/AIE 2016. LNCS (LNAI), vol. 9799, pp. 753–767. Springer, Cham (2016). https://doi.org/10.1007/978-3-319-42007-3_64CrossRefGoogle Scholar