Neural Network-based End-effector Force Estimation for Mobile Manipulator on Simulated Uneven Surfaces

Mobile robotic manipulators often interact with other robots, humans or the environment in indoor and outdoor scenarios. In many cases, end-effector forces need to be known to give feedback about task completion. The mobile base might be titled due to the uneven surface on which the mobile base is positioned. The paper presents the approach to estimating end-effector forces based on neural networks in such cases. The estimates are inferred based on the force sensor mounted under the robot's base and the knowledge of the tilt angle. The robot's dynamic model does not have to be known since it is learned from data during neural network training. The dataset for this research was obtained in simulation. The angle between the robot and the surface changed to simulate a change in surface slope that a mobile manipulator might encounter during the execution of real-world tasks. The trained neural network shows good performance no matter the angle between the base and the ground. It showed an RMSE of 0.302 N (on the test set). Furthermore, there was no significant difference when comparing RMSE across all test data with test data obtained on a per-angle basis, demonstrating the effectiveness of the proposed approach.