Estimation of the legs’ state of a mobile robot based on Long Short-Term Memory network

In this paper, we propose a model-free method for estimating the height and the Ground Reaction Force (GRF) for the legs of mobile robots using the Long Short-Term Memory network (LSTM). The method does not require the presence of a force sensor at each foot, and it is proven to be robust to the changes that may occur in the dynamics. First, we generated a dataset to estimate the state of the legs for the non-damaged robot and for various types of damage situations; a disabled leg with working joints’ encoders, a fully disabled leg, and a removed leg. The network was tuned to obtain the highest stable $R^2$ score. Then, we studied the effect of the available sensors on the results of estimation which proved the sufficiency of using just the joint encoders which led to reducing the computational time by 17%. The sequence length required for estimation is also optimized to less than half of the gait period. The estimation results on a simulated hexapod robot and on a dataset recorded using a real four-legged robot proved the effectiveness and reliability of the proposed method as the $R^2$ score reached 94% with the damaged hexapod robot and 92% with the real four-legged robot, and that also proved the ability of our proposed method to be generalized to different types of robots.