Injured Robot Learns to Walk
Almost every robot has problems with being in balance and moving on an accurate way. Also, making a walking robot could be a complex problem because of the necessary to study the robot specific design and to study the details of the environment which the robot will examine and walk into. Also,study what will happen if the robot get injured (lost one or more of his arms or motors),so the way to solve these problems stars with studying the nature of the animals movements and try the discovered patterns on the manufactured robot. But this solution doesn’t have in mind the differences between the animal muscles and the robot motors, and here came the machine learning approach to solve this problem. This project presents a study on Quadruped with different designs learn how to walk without prior knowledge of explicit dynamics model by using reinforcement learning and reward approach telling whether it’s moving forward. Results show that the robots learn to walk using the presented approach properly. The study tries different components (sensors, micro-controllers) and provide the strong and weak points of every component. As simple as possible the robot will make a random moves and the raspberry pi will save the these moves on array with the readings from the gyroscope (or distance) sensor, so it can know the effect of every motor movement on the robot.