Embodied AI Meta-alignment Workshop @ AAAI-27

Alignment has become one of the central concepts in discussions of embodied AI, but its meaning has become increasingly diffuse. Embodied AI systems are variously described as aligned when they follow instructions, satisfy user preferences, avoid harmful behavior, respect social norms, or act consistently with broader human values. Yet, these objectives raise foundational questions that remain unresolved: aligned with whom, and with what? Are human preferences sufficiently coherent and stable to serve as alignment targets? Can individual preferences be meaningfully aggregated into societal values? Are rewards and value functions appropriate representations of human values? And what should an AI system do when preferences are incomplete, conflicting, context-dependent, or themselves shaped through interaction with the system?

The Embodied AI Meta-alignment workshop treats the dynamic selection of alignment targets and the natural flow of information between system and human as the objects of study. We aim to bring together a community of researchers at the intersection of robotics, AI, and Human-Robot Interaction to study the embodied meta-alignment problem which is concerned with the development of AI agents that dynamically reason about whose preferences to prioritize, which alignment strategies to apply, and how to balance user objectives with safety and norms. We explore how, in physical environments, AI agents have the capability to go beyond satisfying human preferences and through their actions in the environment, can actively influence human trust, expectations, and future preferences. Given a robot’s ability to influence the interaction dynamic and future human state, we explore a coadaptive view of the interaction under which the robot and human interact with and influence each other, to either a positive or negative end. A key outcome of the research, community discussions, and networking that will occur in the workshop is identifying methodologies to ensure the advancement of AI capabilities while understanding and limiting its negative influence.