A fair starting line.
Give every controller the same task, initial state, action limits, and run budget. Compare success, completion time, and invalid actions across repeatable trials.
First candidate: a small, deterministic manipulation task with a recorded observation/action log for every attempt.