Robots learn better when they get small hints at every step instead of only a final thumbs-up or thumbs-down.
Robots often get confused by wordy instructions, so this paper tells them exactly where to touch instead of what to do in sentences.
The paper teaches a game-playing AI to copy good human players (behavior cloning) and shows that simply scaling up the model and the data makes the AI reason more causally (it pays attention to what truly causes outcomes on screen).
NitroGen is a vision-to-action AI that learns to play many video games by watching 40,000 hours of gameplay videos from over 1,000 titles with on-screen controller overlays.