Worlds that answer back
One model imagines the world. One acts in it.
Two models, one loop
One predicts the world. The other predicts the player. Each gets better as the other does.
World model
Takes the current moment and an action, and generates what happens next. No engine and no assets: how the world behaves is learned, not written down.
- Input
- A moment, and an action
- Output
- The next moment
- Worlds
- 2D and 3D
- Status
- In development
Action model
Looks at a world and decides what to press. Small and fast enough to act several times a second, so it can keep up with worlds that do not wait.
- Input
- A moment
- Output
- Stick, buttons, taps and keys
- Pace
- Several actions a second
- Status
- In development
Grounded in play
Both models learn from experience kept the way it happened: what was on screen, what was pressed, and what the world did in return.
Moments and actions, paired
Every frame is kept with what was being pressed at that instant, down to taps too brief to notice.
The world's own account
Next to the pixels, the state the world itself reports, so what a model imagines can be checked against what happened.
Flat and deep
Top-down arcades and 3D runners alike, at the resolution people actually see.
More than one player
Worlds are shared, so the record includes other people acting at the same time.
Good play, not just play
Each world carries a measure of progress, so the action model can learn from what worked.
People and agents
People play when they feel like it. Agents play around the clock, to cover what people skip.
What it has to do
A generated world is worth stepping into only if it passes three tests.
Still early
Leave your email and we will write when there is something to step into.
Questions, or want to work on it? Write to founders@pogl.com.