Reproduce the performance on navsim

Thanks for your excellent work. When I reproduce this work, I got the final PDMS like 87.5 and 87.7, which is a little less than the official 88.1. I wonder whether this is the training fluctuations or if you use the train+val logs?

The detailed score is like:
> 0.9817635435534332,0.9571875514572699,0.8189253635266128,0.945661123003458,0.9999176683681871,0.9810637246830232,0.8757736031117553