Fi0

Fi0 separates reusable task and world knowledge from the actions required by a particular robot body. Its embodiment representation describes kinematic topology, geometry, sensing, end effectors, actuation and dynamic state, allowing a shared model to reorganize execution as those properties change. Cross-view learning connects human first-person demonstrations with robot observations without requiring pixel-level correspondence.

The model combines an Embodiment Graph, Cross-view World Encoder, Skill Encoder and World Dynamics Model. A human demonstration can introduce an unfamiliar task during inference, while the dynamics layer predicts the embodiment-conditioned outcome of candidate actions before execution. Feagine is initially studying the approach on single-arm, dual-arm and multi-arm soft robots.