I’m planning to go on a robot football competition where two robots are placed in an arena and have to collect and score the most ping pong balls into the opponent’s goal.
I’ve managed to build a custom classifier based on the Keras ver. of You Only Look Once that can run in reasonable time on a small computer like the Raspberry Pi. I can train for classes such as “ball”, “goal”, “corner” etc. My problem is that I don’t know how the enemy robots will look like. All I know is the maximum allowed dimensions. They might have visible wheels, or not. They might show that they contain balls, or not.
How should I go about training for such a class if the real-world situation might have a robot very different from what has been seen before?