We have seen advances in top down, RTS team games like Dota 2 and Starcraft II from companies like OpenAI who developed agents to beat real pro players most of the time. How would similar learning techniques compare to games like Overwatch that require faster reaction times and complex understanding of 3d space and effect? Or have we not developed solutions that could be tasked with this problem?
You have mentioned it in the question. reaction time
Even though both the games require strategy and understanding of the surroundings and team play, FPS games are heavily dependent on
reaction time and
accuracy. As much as you try, it's next to impossible to reach the reaction times and accuracy of computer agent.
In games such as DOTA2, the game is more dependent towards
strategy and how the characters interact with each other (Strengths and Weakness).
Finally, it doesn't make sense to train an AI based agent to play FPS games, because the optimal solution is already present and there is no way that AI can reach those levels. It's called AIM BOT. As good as you can get at strategy and planning, team work and reaction times, you can almost never beat it. It's next to 100% accurate and the opponent is basically dead as soon as he's in sight.