TinyWorlds DreamDojo is a personal research sandbox for action-conditioned video world models.
The system is built around a PyTorch pipeline for:
- FSQ video discretization
- latent action encoding
- a compact 19M-parameter dynamics model
- rollout-based planning experiments
- fixed-split evaluation against learned-video, copy-last, and patch-flow baselines
The main empirical thread is high-motion behavior. A training/inference mismatch in future-frame decoding reduced forecast quality on hard examples. After fixing that mismatch, the H1 high-motion oracle rate rose from 30.3% to 63.6% against the strongest baseline, with follow-up decoded-frame loss and action-conditioning ablations.