Desenvolvimento de técnicas de estabilização de caminhada para robô humanoide com detecção de diferentes tipos de terrenos

Abstract

Humanoid robotics has been motivated by possibilities of using bipedal robots to assist humans in domestic tasks, entertainment and in hostile environments to humans, such as natural or human-induced disasters. To deal with the most varied types of environments, it is very important that the robots can develop adaptable locomotion strategies according to its surroundings. In this context, this work presents humanoid locomotion techniques, introducing a stabilization strategy for bipedal locomotion and implementing a new approach for terrain classification during walking. For the walking trajectories generation, the linear inverted pendulum method was used, and this method describes the robot’s dynamics as a simple inverted pendulum. This approach demonstrated the ability of providing stable trajectories for a simulated model of the humanoid robot. On the other hand, some adaptations in the method was made for the real robot, such as adjusting the robot posture and changing the period of each step. Afterwards, a study was made on the impedance control, where it was noticed that changing the stiffness constant of the PID control at joint level has the ability to improve the walking stability of the robot, even in different types of terrains. Finally, a new terrain detection technique was proposed, where a gray scale picture is used as a terrain signature, and this image is generated by reading the data of torque sensors and the inertial sensor at the impact moment. A convolutional neural network is used to classify six different types of terrain using the signature picture as an input, and the overall accuracy of this method reached values above 96.0%, which is comparable to state of art approaches.