Share it

Robotic systems need to operate in open and complex environments where complete knowledge of the world is impossible. This talk explores how Multimodal Large Language Models ((M)LLMs) can support robotic reasoning under conditions of uncertainty and also addresses the challenges posed by uncontrolled actors such as humans.

We will push beyond the direct application of MLLMs by introducing techniques to control hallucinations and ensure robust reasoning.

These techniques include integrating formal methodologies and neurosymbolic approaches to craft more reliable and adaptive reasoning processes for real-world robotic systems.

Programa

  • Introduction: The Problem of Open Environments in Robotics
  • (M)LLMs as Universal Sources of Knowledge
  • Controlling Hallucinations and Enhancing Robustness
  • Grounding Perception and Learning from Robot Data
  • Flexible Plan Generation based on Natural Language
  • Uncertainty Management and Human Preferences
  • Detection and Resolution of Failures in the Execution of Actions
  • Discussion: The Way Forward in Robot Reasoning in Open Environments

Masterclass taught by Eurecat

Taught by:

Magí Dalmau Moreno
Head of Cognitive Robotics at Eurecat

Magí Dalmau earned his degree in Industrial Engineering from the Polytechnic University of Catalonia in 2019 and completed his master’s degree in Artificial Intelligence at Pompeu Fabra University in 2022.

He joined Eurecat in 2019 as a researcher in the Robotic Handling group. Since January 2024, he has been the head of the Cognitive Robotics group, leading research to equip robots with natural reasoning and interaction capabilities in human-centric scenarios leveraging Generative Artificial Intelligence, Reinforcement Learning and AI Planning.