We introduce AtomEgo, a systematic study of ego‑robot co‑training for embodied foundation model pretraining. The work is built on a curated corpus of roughly 2,659 hours and a scalable data‑processing pipeline. For vision‑language‑action and world‑action model architectures, we explore three representative paradigms: joint co‑training with domain‑specific action heads, progressive ego‑to‑robot transfer via embodiment alignment, and joint video‑action modeling.
The paradigms are evaluated through multi‑task real‑robot experiments and language‑conditioned cross‑embodiment representation analysis. Our findings reveal a simple principle: $$Data\ Scale \times Alignment\ Quality \rightarrow Capability\ Gain$$ Egocentric data can improve generalization, but their benefit hinges on how well they are aligned and utilized. This principle offers practical guidance for scalable ego‑robot pretraining.
Review