Each seed runs MAP-Elites and a plain (μ+λ) EA with the same budget per iteration, the same mutation and the same re-check rule (would-be elites scored by the worst of N runs). The EA selects on distance alone; every robot it evaluates is also recorded into a map so both can be scored on the same cells. At the end every elite is re-run from two fresh jittered starts and scored by the average, so luck during the search does not count. Scores in each cell are divided by the best any run found there.

Results

Progress (mean over seeds)

Maps (latest run of each method)