Transfer Learning for Collision Localization in Collaborative Robotics

Transfer Learning for Collision Localization in Collaborative Robotics

연구 분야: Artificial Intelligence

학회: APPIS 2020: Proceedings of the 3rd International Conference on Applications of Intelligent Systems

In this work, the transfer learning paradigm is used to improve the performance of neural networks. These networks estimate the collision point between the robot and an external object, using two different approaches. One of them uses a neural network to predict collision point on the internal robot axis and the second uses classification in order to find a collision in points which are sampled on the robot surface. The neural networks are trained on two datasets, one dataset is generated in simulation, the other one captured from the real robot. Obtained results show, that using a pre-trained network allows to greatly increase the overall accuracy of collision localization.