Adaptive Random Search based Evolutionary Learning of a Humanoid Robot

In this thesis, an on-line learning system of humanoid robot has been developed for robot motion pattern modification. There are a lot of environment and robot uncertainties when humanoid robot moving, this learning system could find out the modification of motion pattern to overcome all uncertainties through the computation of sensor-motor relation.
The learning process is based on adaptive random search (ARS) with reinforcement learning. Sensor signals of motion are used to calculate the fitness function for reinforcement learning. There are one two-axis accelerometer and two one-axis gyros on the head of the robot and four pressure sensors on the feet.
At present, initial gait pattern always makes the robot falls down. After learning process, the stable gait pattern was found. Further, to find out the fastest pattern, the different gait patterns and speeds were composed. In this case, the robot maximum walk velocity is 1386.207mm/min, and could walk on a ramp with 2.85 degree of slope.
This system is uncomplicated, easy to adjust for different conditions and deal with all uncertainties at once. User only need to give a rough initial pattern and the suitable fitness function, don’t need the exact simulation. It could apply on many fields and embedding in robot.