I've been collecting data on CS2 case farming bots for months, training an ML model to detect their movement so I can capture clips and create a mega database on all the accounts.
This worked surprisingly well as with the correct features, their movements was fairly easy to classify.
I managed to get an precision of 98% and accuracy of 94%.
Biggest factor in gathering this much labelled data was auto-labelling based on if they stopped moving when I entered spectator, which the bots typically do to avoid being seen, but in this case it worked against them.
Looking forward to returning to this project for V2 and improving on it, hopefully with a group effort.