Skip to Main Content
We made a flying object with 4 CMOS video cameras and 4 propellers. The processor consists of a visual processing module, attitude calculation module and adaptive module implemented by using neural feedback network. The visual processing modules, which are including 3 CMOS cameras, output displacements of images for each local area using the difference between frames. The attitude calculation module integrates those displacement values and output rotational velocity of each axis (yaw, roll, pitch). The last CMOS camera observes the ground, and its processing module output relative position between the ground and the flying object. The adaptive module finally calculates the motor outputs basically depending on PID feedback control loop with reinforcement learning, and the learning method is weight changes as neural network between derived parameters. A preliminary manual tuning fixes initial parameters of the PID control. An online learning could tune those parameters more precisely during actual fright experiments. We introduce nonlinear response functions and a switching mechanism of those responses into the adaptive module. For instance, nonlinear response functions are as follows: 1) reduce feedback gains to cool down oscillation, 2) add offset value to cancel supposed disturbance, and 3) top all motors to avoid fatal damage.