From-scratch PyTorch reproduction of DreamerV4 (Hafner et al., 2025): masked-autoencoder tokenizer, block-causal flow-matching dynamics with bootstrap curriculum, agent-token finetuning, and PMPO imagination RL. Trains end-to-end on a single GPU
reinforcement-learning pytorch transformer model-based-rl dreamer world-model flow-matching dreamer-v4 pmpo
-
Updated
Jun 24, 2026 - Python