Touchdown control in bipedal robot walking

Abstract

Currently a lot of research is performed in the field of walking robots and especially on bipedal humanoid robots. Successful walking robots have been made but there are still problems that have to be overcome to make them as versatile, robust and energy efficient as human beings. One of these problems is the control of balance after taking a step during which often instability occurs which can result in a fall. To make TUlip step more reliable the cause of this instability has to be determined and solved. To find the cause of this instability push recovery experiments have been performed on the bipedal humanoid walking robot TUlip. After taking a step the upper body starts performing oscillating translational motions and its feet start tilting which cause TUlip to become unstable resulting in a fall. By analyzing the whole control loop layer by layer it was found that the application of too high ankle torques is the cause of this problem. These too high ankle torques are caused by an incorrect control model of the ankle torques and by incorrect desired ankle torques. The desired ankle torques are incorrect due to an incorrect estimation of the Center of Mass (CoM) and an incorrect estimation of the support polygon. The CoM is used to calculate the Instantaneous Capture Point (ICP) which is a point used in the control of balance and to calculate step locations. The ICP and the support polygon determine the location of the desired Center of Pressure (CoP) which the robot uses to determine its desired ankle torques. Since the CoM and support polygon estimation are incorrect this will automatically result in incorrect desired ankle torques and incorrect step locations. So to improve the balance control the CoM estimation, the support polygon estimation and the ankle torque control model have to be improved. Since improving the CoM estimation is already performed by Bart Vissers [1] and improving the support polygon estimation requires hardware improvements, this thesis focuses on improving the ankle torque actuation model. A new ankle torque actuation model is derived and implemented on TUlip. After tuning the controllers the old and the new situation are compared using push recovery experiments. It was found that using the new ankle torque actuation model the number of ground contact changes for the stance foot after taking a step has been reduced with 56.8 %, the distance traveled by the upper body in x-direction after taking a step has been reduced with 37.2 % and most importantly the percentage of successful steps has increased by a factor four.