Testing neural network crash avoidance systems in mobile robot

The ability to observe the movements of obstacles is essential for a robot to safely navigate in a dynamic environment. We use a biologically inspired approach to avoid moving obstacles. Simulations show that the well trained crash avoidance system (CAS) can result in reflexive crash avoidance behavior in an environment without a priori information.; For the problem of controlling the movement of a vehicle in an unknown environment where only partial or no information is available beforehand, the vehicle must be equipped with sensors to detect obstacles and to avoid the crash. ROACH is a small robot car with twelve sonar sensors that was developed to test a CAS in real world situations.; Fixed-Point arithmetic is used since the DSP chip (ADSP-2181), a 16 bit fixed point chip, in ROACH doesn't have enough memory space for the CAS program to be implemented in floating point. Fixed-Point arithmetic is the method that changes floating point math to fixed point without losing significant accuracy. Fixed-Point is less accurate but reduces the program memory tremendously and the speed of computation is much faster.; The fixed point CAS program is converted to Assembler for implementation in the DSP chip which controls the vehicle. The final program and data are downloaded to the DSP chip and the vehicle is run. The sonar sensors measure obstacle data, distance and angle from the vehicle-fixed reference frame at each interval time and send them to the DSP chip. DSP chip calculates vehicle steering angle and speed, which are needed for the vehicle to avoid obstacles.; Simulation is used to train and demonstrate the CAS neural network. The CAS is shown to be effective more than 70% of the time for a dynamic environment involving one or more moving obstacles. ROACH with the CAS installed is shown to successfully avoid static obstacles and some moving obstacles.