ANN-UKF-based estimator for landing forces in quadruped robots

Estimation of landing forces is essential for legged robots to execute precise and effective movements, particularly when navigating complex and unfamiliar terrains. Traditionally, this estimation is achieved using force sensors mounted on the robot's legs. This paper introduces a novel approach that eliminates the need for force sensors by employing a force-sensorless estimator. The proposed method utilizes an artificial neural network (ANN) to forecast joint torques at the legs at specific points along the motion trajectory. This ANN model functions as a predictor within an Unscented Kalman Filter (UKF), which refines and stabilizes the force measurements. By analyzing the differences between predicted and filtered values, the method estimates both the state and magnitude of the robot’s foot impact with the terrain. The effectiveness of this approach has been validated through simulations and practical tests on a quadruped robot.