Enhancing DCM-based Walking Using a Generalized eCMP Adaption for Compensating Modelling and Measurement Errors

Abstract

In the future, robots shall be used in many areas to simplify human lives. Human-like robots, which can move bipedally, could be integrated into the society. Most importantly, they will enable missions that can be risky or dangerous for humans. Therefore, the area of bipedal walking for robots is currently a large and popular field of research. Various approaches are being developed to model, plan, and control the complex and high-dimensional motions of robots. Many of these approaches are based on the dynamics of the robot's center of mass. Recently, a new gait framework has been introduced for this purpose which is based on the Divergent Component of Motion. This focuses on the critical unstable part of the center of mass dynamics. In some previous work, corresponding successes have been achieved in terms of stable robust walking with this framework. It has also been used for special situations such as uneven ground or balancing scenarios. Two methods in particular have proven successful. The use of Iterative Learning Control as a learning algorithm as well as taking into account the change in total angular momentum. The results obtained were still limited in performance due to measurement and modeling errors. Therefore, in this work a method was developed which merges both variants and thus a significantly improved stable walking of the humanoid TORO at DLR could be achieved. Furthermore, a database with numerous recorded gait trajectories for different walking parameters was created in simulation using the new method. This was then used to develop a neural network that predicts the resulting trajectories of the converged algorithm without required feedback simply from the information of the desired gait trajectories. The neural network was then implemented in the simulation. The predicted trajectories were quite similar to those from the recordings and the overall gait stability was significantly improved compared to the reference case without learning