Orbis 2 presents a hierarchical world model for driving that separates future prediction into a high-level predictor for coarse scene structure over longer horizons and a low-level generator for detailed outputs. The authors combine diffusion forcing during pretraining with teacher forcing during fine-tuning: the former is reported to produce richer internal representations, while the latter provides more stable autoregressive rollouts. The paper reports state-of-the-art performance across established driving-world-model evaluations, including long-horizon generation fidelity, counterfactual steering responsiveness, and representation quality. Code, demos, checkpoints, and qualitative results are listed on the project page.
No heat snapshots are available in the last 24 hours.