ZIPDROP neon courier
SLOW
careful corridor
00:00.00
0 steps
FLYING

Race the laws, same seed

Keep the drops coming.

Race Slow against Fast on one seed — that is the show. Fly it yourself with a watch AI beside you; miss and you are back in the air on the same seed.

  • W↑ thrust
  • AD strafe
  • S↓ sink
  • R restart

Research demo: world model → land → go fast —

Research example

What this demo is about

ZipDrop is a tiny playable sketch of a control recipe we study in ParticleGAN — a PyTorch library for GANs with learnable particle priors (a finite set of latent vectors optimized with the generator, plus the usual relativistic / cap-regularized critic tools). You do not need the library to play this page.

The research pipeline has three stages. This browser game only illustrates the slow vs fast race at the end.

1. Learn a world model with GANs

First, train a transition model on logged simulator steps: predict state, action, and next state with generators and critics (plus a particle prior). That gives you a generative model of “what the world does,” not a policy that scores points yet.

2. Learn something that lands

Freeze most of that world model. Finetune a small controller head so it copies successful expert actions (and/or a paired adversarial controller objective). Ranking is landings first. A strong result on the lander board was 50/50 successful landings after this stage — the world model alone does not “land.”

3. Learn something that is fast

Keep only successful rollouts. On the same start (same seed), pair a careful success with a faster success. Finetune toward the faster actions with a paired adversarial step; crashes never count as “fast.” The student should cut time among successes without collapsing into crashy shortcuts.

What you are playing

Race the AIs puts a careful law and a faster law on one seeded route so you can see stage 3’s tradeoff live. The pilots here are simple deterministic laws for the demo — not a live ParticleGAN train in the browser. The repo links are the real training stack.