Data Philosophy
Our approach is rooted in a fundamental observation about embodied intelligence: data determines generalization. In high-dimensional, continuous robotic control spaces, the scarcity of real-world interaction data severely limits model expressivity and transferability. Unlike language or vision models trained on internet-scale data, embodied agents face a "data desert"—sparse, expensive to collect, and often non-i.i.d.
This leads to a critical bottleneck: without sufficient coverage of state-action trajectories, even powerful diffusion models fail to generalize beyond training distributions. We argue that compositionality through primitives is the key to breaking this barrier— enabling rich behavior from limited data by recombining learned skills.
Figure: PrimitiveWorld achieves higher data efficiency by decomposing complex tasks into reusable action primitives, reducing the need for exhaustive task-specific data collection.
By modeling short-horizon dynamics within primitive-conditioned world models, we decouple skill learning from planning. This enables zero-shot composition— executing unseen task combinations by reusing learned motion priors. As shown below, this architectural choice unlocks strong generalization even in low-data regimes.
Figure: Generalization in Diffusion Models. PrimitiveWorld outperforms end-to-end approaches on unseen (predicate, object) combinations, demonstrating compositional reasoning enabled by structured world modeling.
“Structure your model where data is scarce, and let it learn where data is abundant.”
.gif)
.gif)
.gif)
.gif)
.gif)
.gif)
