Autonomous Learning of Bipedal Walking Stabilization
Final Report Abstract
Humanoid robots, robots with a human-like body plan, enjoy increasing popularity as research platform, because they have the potential to be universal robots that can perform a large variety of tasks in environments that are designed to suit human needs. It is, however, a big challenge to replicate efficiency, robustness, and grace of the natural human gait. Bipedal walkers are inherently unstable and difficult to control. In the first funding period of the Priority Programme Autonomous Learning, we developed a novel closed-loop control approach for bipedal walking in the presence of large disturbances. Our capture step controller is able to absorb disturbances that can occur from any direction and at any time during a step and returns the robot to the desired walking velocity within one or two steps. To make the first capture step even more effective, we developed a method to learn in an online fashion a non-parametric model describing deviations from the point-mass abstraction from a small number of observations of residual energy after taking capture steps. In this project, we transferred the capture step framework to three new humanoid robots: igus Humanoid Open Platform, NimbRo-OP2, and NimbRo-OP2X. The AdultSize NimbRo-OP2(X) robots have been developed in this project. They are 135 cm tall, have 3D printed structural parts with a parallel kinematic mechanism in the legs, and feature strong onboard computers with GPUs. Hardware and software of these robots has been released. In this project, we investigated more advanced push recovery strategies that went beyond capture steps, i.e., using the rotational inertia of the torso and the arms to complement capture steps with balance-restoring torque. The balance of our robots has been advanced by fused angle based feedback mechanisms, tilt phase feedback, and the real-time generation of dynamic full-body motion based on a five mass model. Parameters of balance controllers were learned online from few real-robot and a larger number of simulated robot experiments by Bayesian optimization. Furthermore, we developed methods for deep-learning based scene perception and for online adaptation of steps to visually perceived constraints on foot placement. We integrated the developed methods in our humanoid soccer system. The disturbance rejection was tested in isolation in the Push Recovery technical challenge, where our robots regularly achieved top scores. We also used the developed capture step, fused angle and tilt phase feedback mechanisms in the games, where they greatly contributed to the stability of our robots. The robots of our team NimbRo won all international RoboCup Humanoid League competitions where they participated during the project duration.
Publications
-
(2017). NimbRo-OP2: Grown-up 3D printed open humanoid platform for research. In IEEE-RAS 17th International Conference on Humanoid Robots (Humanoids)
Ficht, G., Allgeuer, P., Farazi, H., and Behnke, S.
-
(2018). Combining simulations and realrobot experiments for Bayesian optimization of bipedal gait stabilization. In 22nd RoboCup International Symposium
Rodriguez, D., Brandenburger, A., and Behnke, S.
-
(2019). Bipedal walking with corrective actions in the tilt phase space. In IEEE-RAS 18th International Conference on Humanoid Robots (Humanoids)
Allgeuer, P. and Behnke, S.
-
(2019). Capture steps: Robust walking for humanoid robots. International Journal of Humanoid Robotics (IJHR), 16(6):1950032:1–28
Missura, M., Bennewitz, M., and Behnke, S.
-
(2019). Online balanced motion generation for humanoid robots. In IEEE-RAS 18th International Conference on Humanoid Robots (Humanoids)
Ficht, G. and Behnke, S.
-
(2020). Fast whole-body motion control of humanoid robots with inertia constraints. In IEEE International Conference on Robotics and Automation (ICRA)
Ficht, G. and Behnke, S.
