NVIDIA and KAIST Researchers Say BirdDreams Cuts Trajectory Error 35 Percent
Summary
NVIDIA and KAIST researchers present BirdDreams, a driving world model that jointly generates perspective video and a bird's-eye-view video from a single camera image, without HD maps or surrounding-vehicle states at inference. The team reports 35% lower trajectory error than an OmniDreams-based baseline on PhysicalAI-AV, 3.83 meters versus 5.88.
Key Points
- Researchers report a 21% lower trajectory error without ego-action input, 6.15 meters minADE against 7.76 for the baseline.
- Joint attention links two video diffusion transformers, one per stream, so the perspective and BEV views inform each other, the team says.
- In a 375-judgment human study, 25 participants pick BirdDreams over the baseline 57% of the time for realistic traffic behavior.