Nonlinear State Estimation for a Bipedal Robot

Abstract

Humanoid robots could perform a large variety of tasks, ranging from helping the elderly and entertainment to carrying out repairs in hazardous environments such as nuclear power plants. Therefore, they are an important topic of research. However, controlling the balance of two-legged robots still proves to be a challenge, especially during more complex motions such as walking. One of the issues is inaccurate knowledge about the state of the robot. Specifically, the Center of Mass and the locations of the feet are required for balancing control. Because of uncertainties in the kinematic chain such as play, link flexibility and parameter uncertainty as well as sensory noise, they are often difficult to obtain. Therefore, a state estimator is proposed making use of the robot’s full-body dynamics to predict the next state. The Unscented Kalman Filter is chosen to fuse the predictions with the measurement. To evaluate the new filter design, a second estimator is implemented as a benchmark, using a conventional Linear Inverted Pendulum Model for the prediction and a Kalman Filter for data fusion. Simulations and a preliminary application were used to investigate the robustness and accuracy. Estimation with a pendulum model gave incorrect estimates in the presence of sensory bias. Using the full-body dynamics, however, the effect of sensory bias was reduced significantly. Moreover, the proposed method was shown to be robust against parametric errors. However, the performance varied between different motions, making it hard to tune. Improving the filter to inherently work for various movements as well as making the filter numerically efficient enough for online implementation still requires further research.