I wonder what the input for the bot was. Just a screengrab like in other reinforcement learning examples (like Pong or GTA etc.) or did they use some Dota API t

I wonder what the input for the bot was. Just a screengrab like in other reinforcement learning examples (like Pong or GTA etc.) or did they use some Dota API to make their lives a bit more easy? Impressive nonetheless, like others said, knowing the reward function would be interesting.