This is a World Model playing Flappy Bird — but there is no game running. Everything below is a “dream”: a DreamerV3 world model hallucinating each frame from its own imagination. A learned actor flies the bird inside that dream. Press Space (or tap the frame) and you take the stick.

How it works

Unlike the two-stage World Models recipe (a VAE, then a separate memory RNN), DreamerV3 learns one joint world model end to end — a Recurrent State-Space Model (RSSM). Its latent state at each step splits in two: a deterministic part h (a GRU carrying history) and a stochastic part z — here 32 independent 32-way categoricals, sampled fresh each step. Everything below runs frozen and exported to ONNX, entirely in your browser with onnxruntime-web.

The RSSM (dynamics). A GRU folds the previous stochastic state and your action into h, and a prior network predicts the distribution over the next z from history alone — no observation needed. Rolling that prior forward is the dream. During training a posterior that also sees the real frame is pulled toward the prior (a KL term), so the prior alone becomes a competent forward model of pipes and gravity.

The decoder + heads. From the combined feature [h, z], a convolutional decoder reconstructs the frame you see, while reward and continue heads predict the expected reward and the probability the episode keeps going. That continue probability is what ends a dream when it drops.

The actor. A small policy mapping [h, z] to an action, trained entirely on trajectories the world model imagines — never on the real environment.

Each tick the loop is: decode [h, z] to the frame you see → the actor (or your Space bar) picks flap / no-flap → the RSSM advances its GRU and samples the next z. When the continue head’s probability crosses a threshold, the dream ends and a fresh one is seeded from a real starting state. Your flap and the actor’s share the same stick, so you’re steering the policy through its own hallucination.