Programme time: 14-Oct | 10:00 | 75 min
Room: Arrabida Room 

Moderator: Marta Lagomarsino
Speakers
•    Marta Lagomarsino
•    Valerio Bo

•    Néstor Garcia Hidalgo

Full description and objectives

Foundation models are rapidly changing the way robots perceive, reason, communicate, and act. Yet their deployment in real-world robotics remains challenging, especially when robots must manipulate small, soft, delicate, or deformable objects in dynamic, human-populated environments such as hospitals and factories.

This session will explore how foundation models can support dexterous manipulation, adaptive task planning, semantic perception, human-robot interaction, learning from demonstration, and explainable decision-making. The session will focus on concrete challenges inspired by real use cases in healthcare and industry: changing IV drip sets in a hospital ward, handling delicate gears and deformable ply-sheets in manufacturing, and adapting robot behaviour when objects, users, or operating conditions change unexpectedly.

The objective is to identify practical pathways for moving foundation-model-based robotics from promising prototypes to validated, trustworthy deployment. Participants will discuss technical, operational, ethical, and regulatory barriers, and will co-design validation strategies for robots that continuously learn and adapt in safety-critical environments.

The session aims to:

• Discuss methodologies for validating trustworthy, safe, and human-centric robotic systems under realistic operational conditions.

• Explore the integration of language, perception, reasoning, and learning capabilities to enhance robotic autonomy and adaptability.

• Examine emerging skills and expertise, and workforce needs required to support the next generation of intelligent robotic systems.

Why you want to join?

This session addresses the convergence of AI and robotics through foundation models for adaptive, dexterous, and safe robot interaction and manipulation in real-world environments. It aligns with ADRForum priorities on Europe’s resilience, competitiveness, and strategic autonomy by exploring trustworthy physical AI for strategic sectors such as healthcare and manufacturing.

The session will examine how large language models (LLMs), vision-language models (VLMs), multimodal perception, adaptive cognitive pipelines, learning from demonstration (LfD), and Explainable AI (XAI) can enable robots to operate beyond controlled laboratory environments. Their combination allows robots to understand instructions and context, integrate visual, tactile, force, and environmental information, adapt their behaviour, and acquire new skills from human demonstrations.

Particular attention will be given to real-world deployment challenges, including safety, validation, human trust, domain adaptation, regulatory readiness, and value creation with end-users. Healthcare and industrial production provide demanding scenarios where robots must safely manipulate small, soft, delicate, or deformable objects while interacting with people.

The session will also explore how these technologies can strengthen Europe’s capacity to develop strategic and trustworthy AI and robotics capabilities while reducing technological dependencies. Explainability, continuous adaptation, and human-centred interaction can support certification, supervision, and confidence in robotic systems. Natural-language interaction and demonstration-based learning can further lower technical barriers, enabling healthcare professionals and industrial workers to transfer knowledge to robots with limited programming and supporting responsible adoption, skills development, and economic and societal value.

Running agenda