World models for robotics
3D geometry-aware representations for world models and world action models.
World models predict how a scene changes under an agent’s actions. For robotics, I am interested in world models and World Action Models whose representations respect 3D geometry, so that predictions remain consistent across viewpoints and over time.
Questions I am working on:
- Which inductive biases make latent dynamics geometry-aware without sacrificing scalability?
- How should action tokens interact with visual tokens in a shared representation space?
- How can the representation-geometry view from my Neural Collapse work inform the design and analysis of world models?