Friday, 13 November 2026 · 7 min read
BirdDreams jointly generates perspective and BEV videos for reactive driving simulation
BirdDreams predicts future driving scenes by jointly generating perspective and bird's-eye-view videos, enabling better ego-trajectory control and realistic surrounding-vehicle reactions without requiring HD maps or vehicle states at inference.
Based on the paper BirdDreams: Joint Perspective and BEV Video Generation for a Plausible, Reactive Driving World Model by Junyoung Seo, Jaewoo Seo, Seungryong Kim, Laura Leal-Taixé and 2 others
Challenges in learning driving world models from perspective video alone
Driving world models simulate how a driving scene evolves under different ego vehicle actions, requiring plausible motion of surrounding vehicles and precise ego-trajectory control. Learning this from perspective video alone is difficult because vehicle motion is entangled with appearance, lighting, and occlusions, making it hard to infer true vehicle trajectories and interactions.
Prior methods condition video generation on external layouts specifying vehicle motion, which provides control but requires future trajectories as inputs and cannot model reactive behavior of surrounding vehicles to new ego maneuvers. Without such inputs, models often produce implausible vehicle behavior or disappearing vehicles at complex intersections.
A key challenge is representing scene layout and vehicle motion explicitly to disentangle dynamics from appearance, enabling the model to predict plausible, reactive vehicle behavior and follow ego trajectories accurately without relying on external maps or vehicle states at inference.
Sign in to read the whole paper
You have read the opening. The rest is free with a Bezen account, and no card is needed.
In the rest of this paper
- How the method works, step by step, with a flow chart
- Three key formulas from the paper, explained symbol by symbol
- Three figures from the paper, and how to read them
- How it was tested, what was found, and where it falls short