Researchers propose a modular approach allowing robots to learn new skills while preserving their core identity through Embodied Capability Modules.

New paradigm separates capability evolution from agent identity, allowing robots to continuously improve without destabilizing their core cognitive system

Introduces Embodied Capability Modules (ECMs) as versioned, modular units of functionality that can be independently learned, refined, and combined

Addresses instability problems in current approaches that rely on prompt engineering, policy updates, or structural redesigns of embodied agents

Enables long-lived robotic systems to persistently operate in dynamic environments while maintaining persistent identity across capability upgrades