How do I catch the rare edge cases when testing robots?

You can test most robots well, so your foundation is sound. The edge cases and rare failures get past you because they are rare, which means the ordinary testing that catches ordinary problems is exactly the wrong tool for them. Catching them is a different discipline, and struggling with it is not a gap in your ability but the frontier of it. Rare failures hide in the situations no one thought to try, so the work is to think like the world at its worst. Ask what happens at the limits: the unexpected input, the two unlikely things going wrong at once, the condition just outside what you designed for. Deliberately try to break the robot rather than confirm it works. Every rare failure you can force in testing is one that will not surprise someone later when it matters far more. Keep a growing catalog of the edge cases that got past you, and turn each into a test you never have to be surprised by twice. You will never imagine every one, and honesty about that limit is part of doing this well. But the range of the rare and the strange that you can anticipate widens with every case you chase down. That widening is the craft, and you are already in it.