The Zeroplay project has the goal of creating an AI agent with these properties:
- Can complete (any%) every video game in Wikipedia's List of video games considered the best,
- Can do this with zero-shot learning (not trained on the listed games themselves)
- And can execute on widely available, retail hardware (latest MacBook)
- A model to map controller screenshots to controller inputs
- A program to map a sequence of video screenshots to time series controller inputs
- A dataset of (screenshot,time)->input mappings collected from various games
- Past-masking to generate unsupervised training data (it it now a transformer)?
- Zeroplay: Zero-shot completion of video game canon on retail hardware