Research seminars

Bio: Paolo Robuffo Giordano is a CNRS Senior Research Scientist at IRISA in Rennes, France, where he directs the Rainbow Team common to IRISA and Inria Rennes. His research investigates resilient autonomous systems, focusing on robust planning, uncertainty propagation, multi-robot architectures, and shared control schemes for single and multiple robots. A core objective of his recent work is developing computationally tractable methodologies for injecting an "intrinsic robustness" layer in modern motion generation schemes, including Model Predictive Control and Reinforcement Learning, to safely deploy robots in unstructured environments. Previously, he held research positions at the German Aerospace Center (DLR) and the Max Planck Institute for Biological Cybernetics. He is a recipient of the 2019 Michel Monpetit Award from the French Academy of Sciences, a Distinguished Lecturer for the IEEE RAS Technical Committee on Multi-robot Systems, and is a Senior Editor for the IEEE Transactions on Robotics.

Abstract: As robots transition from controlled labs to unpredictable environments, achieving reliable autonomy in spite of sensor noise, model inaccuracies, and disturbances remains a formidable challenge. However, successful integration of robots into our daily lives depends crucially on their ability to operate safely and reliably despite these uncertainties.

This talk explores our journey in developing computationally tractable methods for real-world robustness using "sensitivity-based" metrics. I will retrace our research progression, starting with mobile robots (UAVs) and manipulator arms. In these domains, we proposed sensitivity-aware offline and online trajectory planning able to explicitly account for model uncertainty to produce intrinsically robust motion plans. I will then discuss recent extensions to more challenging discontinuous, contact-based dynamics, such as in legged locomotion, and collaborative multi-robot missions. A key advantage of our approach lies in its computational tractability: the proposed robustness metrics are amenable for fast (or even real-time) replanning and, crucially, do not require a forced simplification of the robot/environmental model for arriving at a tractable formulation.

Finally, I will discuss our current and future activities on this research journey: moving beyond robust planning to embed these robustness metrics directly into policy learning algorithms. For instance, by integrating principled uncertainty propagation into the training phase, one can synthesize control policies with an "intrinsic robustness layer" against real-world variations, thus significantly mitigating the well-known sim-to-real gap. Our ultimate goal is to merge rigorous robustness metrics with the adaptability of modern robot learning for developing the next generation of motion generation algorithms for modern robots.