Abstract
An essential ability of a robot is to act in its environment by generating motions for locomotion or manipulation. This can be a challenging problem on a robot with high degrees of freedom. Although biped robots have shown drastic improvements with regard to motion skills over the past few years, many approaches for generating motions still require tedious and time-consuming manual calibration due to variances in the hardware or inaccurate sensors and actuators. This research focuses on generating motions for humanoid robots automatically without manual calibration. The first approach discussed uses parameter optimization (e.g. CMA-ES, PSO) to directly optimize the joint angle trajectories for various motions. Using optimization we produce motions for NAO robots in the RoboCup 3D Soccer Simulation League that are far superior to hand tuned motions or stabilize motions generated from noisy motion capture data from a Microsoft Kinect. The second part describes a dynamically generated closed-loop gait for simulated and physical NAOs using a linear inverted pendulum model (LIPM), which keeps the zero moment point (ZMP) within given constraints. This is a common approach, but due to variances in the hardware and environment it still requires manual fine-tuning. Our experiments show that the model errors can be reduced by optimizing parameters of the model using the observed behavior while walking. Improving the model produces better predictions of the robots behavior which yields a more stable walk without requiring manual calibration.