Embodied AI, i.e., the development of technologies capable of interacting with their physical environment, witnessed substantial progress in the last years, mainly thanks to the advent of Large Language Models (LLMs), Multimodal Large Language Models (MLLMs), and other large foundation AI approaches. These breakthroughs have enabled human users to interact with embodied agents (e.g., robots) through natural language, instructing an embodied agent to navigate to a given position, searching for a specific object, or interacting with the object to perform a certain manipulation.

However, while being inherently human-centric, these problems are often addressed without considering the presence of human users and their interaction with the agent throughout the execution of the embodied task. Considering that the long-term goal of Embodied AI is to create a Co-Habitat for humans and agents, Bi-directional communication (i.e., human-to-agent and agent-to-human) emerges as a key challenge that must be addressed to effectively resolve uncertainty during the full cycle of robotic perception, reasoning, planning, action, and verification.

However, the problem of creating an intuitive, effective, trustworthy, and robust human-agent Co-Habitat remains largely unexplored.

This workshop will address the challenges of creating seamless human-agent Co-Habitats, focusing on trustworthiness (when agents don't know), robustness (when agents ask for help), and intuitive collaboration (bi-directional interaction). We aim to explore methods for integrating MLLMs into embodied agents for effective multimodal interactions, identifying and mitigating the risks of model hallucinations, and modeling human error and communication patterns.