EgoX Framework for Cross-Embodiment Robotic Manipulation
November 11, 2026
EgoX enables robots to learn from mixed datasets of humans, real robots, and simulations by using a Modularized Cross-Embodiment Transformer (MXT+). The architecture uses shared transformer trunks with dedicated tokenizers and action experts for different morphology and sensing modalities.
HOW THIS AFFECTS YOU
●
builderYou can train more capable robot policies using heterogeneous data from different hardware setups.
●
researcherThe MXT+ architecture uses modality-specific experts to bridge the gap between different robot embodiments.