Gepetto
Science
Meetings

Environmental monitoring is becoming increasingly important to address biodiversity loss, climate change, and ecosystem degradation. Traditional monitoring approaches are often labor-intensive, expensive, and limited in spatial and temporal coverage. Recent advances in robotics and artificial intelligence offer new opportunities to develop autonomous systems capable of collecting high-quality environmental data in complex and unstructured natural environments. This talk presents the Natural Intelligence approach, developed within the European H2020 Natural Intelligence (NI) project, which draws inspiration from the adaptive capabilities of biological systems to design robots that are resilient, efficient, and able to operate with minimal human intervention. Rather than relying solely on increasingly complex AI models, the NI paradigm combines morphology, embodied intelligence, and learning to create robotic systems that naturally adapt to challenging environments. The presentation will illustrate how legged robots equipped with multimodal sensing can autonomously navigate forests, wetlands, and protected natural areas, acquiring environmental data while minimizing disturbance to ecosystems. Examples from real-world field deployments will demonstrate how adaptive locomotion, autonomous mission planning, and AI-based perception enable long-term biodiversity monitoring, habitat assessment, and ecological data collection. The talk concludes by discussing how Natural Intelligence can contribute to the next generation of environmental robotics, supporting scientists, conservation agencies, and public authorities in building scalable, sustainable, and data-driven ecosystem monitoring solutions.

Manolo Garabini graduated in Mechanical Engineering and received a Ph.D. degree in Robotics from the University of Pisa, where he is currently an Associate Professor. His main research interests are in the design, planning, and control of soft and adaptive robots, from actuators to end-effectors such as hands, grippers, and feet, as well as complex multi-degree-of-freedom systems. Part of his research has focused on theoretically demonstrating the effectiveness of soft and adaptive robots in high-performance, energy-efficient, and resilient tasks through analytical and numerical optimization methods. He contributed to the development of modular Variable Stiffness Actuators and to the design of the joints and lower body of the humanoid robot WALK-MAN. He participated in the DARPA Robotics Challenge and in a field deployment in Amatrice, Italy, following the 2016 earthquake. He also co-designed the hardware and software of WRAPP-up, a dual-arm robot for flexible picking and palletizing, as well as a robotic system for film unwrapping. More recently, he has contributed to algorithms based on optimal control and learning for minimum-time motion planning and control under uncertainty. He was Principal Investigator for the University of Pisa in the H2020 THING project, coordinator of the H2020 Natural Intelligence project, and coordinator of the Dysturbance sub-project within the H2020 EUROBENCH project. He is currently the Principal Investigator of the FISA-2023 OCCAM personal grant and co-founder of qbrobotics s.r.l. and XStar Motion s.r.l.

Fluid environments present unique opportunities and challenges for robotic systems. This talk explores approaches to designing robots that leverage fluid–structure interactions to generate intelligent behavior. Drawing inspiration from natural systems as well as computational design methodologies, a number of robotic platforms whose morphology – their body shape, material, and structure – plays an active role in sensing, control, and locomotion will be introduced. From sea-creature-inspired robots to “robot scientists” that autonomously fabricate, experiment, and analyze paper airplane flights, this talk will explore how robots, particularly their morphology—their body shape and structure—can be designed to leverage fluid-solid interactions, enabling intelligent behaviors that enhance robustness and efficiency. These design methodologies can play a key role in improving the processes of robot design, enabling the creation of better robots for the real world.

Nana Obayashi is an Assistant Professor at New York University where she established the Prema Lab in 2026. The lab focuses on developing new methods and approaches for designing robotic systems that exploit embodiment, morphology, and environmental interactions. This includes research on bio-inspired robots, robotic automation, and computational design methodologies. She completed her BS and MS in aerospace engineering at Georgia Institute of Technology, USA. Her PhD studies were in robot design and optimization at EPFL in Switzerland, including research visits at the University of Cambridge.

Humanoid robotics is progressing rapidly, but the design, reproduction, and experimental validation of humanoid platforms remain difficult and costly. In this presentation, I will introduce my work at Hugging Face on LeRobot Humanoid, an open, low-cost, mostly 3D-printed bipedal robot designed to make humanoid robot learning more accessible and reproducible.

The project combines mechanical design, simulation, real-world control, and robot learning into a unified development loop. A central aspect of the work is the use of co-design methods, where morphology, actuation choices, simulation models, and control strategies are considered together rather than as separate problems. This approach makes it possible to rapidly iterate on the robot’s structure, evaluate design trade-offs, and connect hardware decisions to downstream learning performance.

I will present the motivation behind building an open humanoid platform, the design principles of the robot, its 3D-printed and off-the-shelf hardware architecture, and the software stack used for simulation, identification, training, and real-world control. I will also discuss the broader goal of LeRobot Humanoid: enabling the robotics community to build, modify, simulate, train, and improve humanoid robots using shared open tools and reproducible workflows.

Virgile Batto is a robotics researcher and engineer working on humanoid robotics at Hugging Face, within the LeRobot team. He recently completed a PhD at LAAS-CNRS / Gepetto, focused on the systematic design and co-design of bipedal robots. His work combines mechanical design, closed-loop kinematics, simulation, optimal control, evolutionary optimization, and hardware prototyping.

At Hugging Face, he works on LeRobot Humanoid, an open, low-cost, mostly 3D-printed humanoid robot platform for robot learning. His broader goal is to make humanoid robotics more reproducible and accessible by connecting hardware design, simulation, learning, and real-world control in an open-source workflow.

As robots move out of the lab and factory and into more challenging environments, uncertainty in the robot's state, dynamics, and contact conditions becomes a fact of life. Uncertainty in contact conditions is particularly challenging because of the discontinuous nature of contact. This talk presents recent work in handling uncertainty in dynamics and contact conditions, both to reduce uncertainty where possible and to generate strategies that do not require perfect knowledge of the world state. Topics include learning off-road driving in novel environments, walking control for legged robots through vegetation, and modeling uncertainty in hybrid dynamical systems.

Aaron M. Johnson is a Professor of Mechanical Engineering at Carnegie Mellon University, with additional appointments in the Robotics Institute and Electrical & Computer Engineering departments. He received his Ph.D. from the University of Pennsylvania in 2014. His research interests include hybrid systems, state estimation, control, learning, legged robots, and field robotics. He is an ASME Fellow and the recipient of the NSF CAREER Award, the ARO Young Investigator Award, and the Best Paper Award at the ICRA Workshop on Legged Robots, among other awards. For the 2025-2026 school year, he is a Visiting Professor and Hans Fischer Fellow at Technical University of Munich.

Multi-robot and human-robot cooperative tasks represent challenging frontiers in robotics. This talk presents recent and ongoing research at the University of Sherbrooke on collaborative autonomy, human-robot interaction, and learning-based locomotion. Topics include autonomous navigation and control strategies for cooperative manipulation, intuitive and physically grounded human-robot interaction, continuous multimodal intention and attention recognition, collision-free motion with diffusion-based models, and learning for quadruped locomotion and loco-manipulation in challenging environments. Experimental results include platforms such as NAO, HRP-4, Unitree Go2, and Baxter.

Waël Suleiman received the Master's and Ph.D. degrees in automatic control from Paul Sabatier University, Université Toulouse III, in 2004 and 2008. His doctoral research at LAAS-CNRS focused on motion optimization for the HRP-2 humanoid robot under the supervision of Jean-Paul Laumond and André Monin. From 2008 to 2010, he was a JSPS Postdoctoral Fellow at the Joint Robotics Laboratory in Tsukuba, Japan. He joined the University of Sherbrooke in 2011, where he is now a Full Professor in the Department of Electrical and Computer Engineering. His research interests span humanoid and collaborative robotics, motion planning and control, human-robot interaction, and learning-based approaches for perception and locomotion in robotics.

Despite significant advances in AI, robots still struggle with tasks involving physical interaction. Learning manipulation skills is hard and fascinating because the movements and behaviors to acquire are tightly connected to the physical world and embodied intelligence. This talk presents representations and learning approaches to help robots acquire manipulation skills by imitation and self-refinement, with an emphasis on frugal learning: learning manipulation skills from only a few demonstrations or exploration trials, and learning only the components of the skill that really need to be learned. The talk will emphasize geometry, manifolds, implicit shape representations, distance fields, and ergodic control as tools for human-guided manipulation skill acquisition.

Dr. Sylvain Calinon is a Senior Research Scientist at the Idiap Research Institute and a Lecturer at the École Polytechnique Fédérale de Lausanne (EPFL). He heads the Robot Learning & Interaction group at Idiap, with expertise in human-robot collaboration, robot learning from demonstration, geometric representations, and optimal control. The approaches developed in his group apply to a wide range of manipulation skills, with robots that are close to us, parts of us, or far away from us.

This talk is about the use of robots to enable closed-loop control of frontal polymerization, an autocatalytic curing process that enables fast and efficient manufacturing of polymers and composites. The approach has two parts: a real-time algorithm that detects fronts from streams of thermal images, and a real-time algorithm that chooses where to apply thermal inputs with a moving ultraviolet source carried by a robotic manipulator. The goal is to steer detected fronts so that they converge to a target profile, meaning a desired shape and speed of propagation. Proof-of-concept hardware experiments show that closed-loop control of front propagation allows target profiles that differ significantly from uncontrolled profiles, despite various sources of uncertainty.

Timothy Bretl is a Professor, a Severns Faculty Scholar, and the Associate Head for Undergraduate Programs in the Department of Aerospace Engineering at the University of Illinois at Urbana-Champaign. He holds an affiliate appointment in the Coordinated Science Laboratory, where he leads a research group working on robotics and education. He has also received every award for undergraduate teaching granted by his department, college, and campus.

The overlapping terminology between the robotics and machine learning communities sometimes creates misunderstandings. Model-free reinforcement learning, for instance, is widely used to train motion policies in simulation and appears to contradict classic model-based approaches. In reality, the simulator itself provides an implicit model. In this talk, we ask what knowledge from robotics science can be formalized as inductive biases for robot learning. We explore this question through the lenses of state estimation and predictive control, focusing on contact detection and obstacle avoidance. By combining insights from robotics with real-world data, we will discuss inductive biases that accelerate training while allowing deployment on real systems.

Stéphane is a researcher in robotics. He received the M.Sc. in computer science from the École Normale Supérieure in 2012, and the Ph.D. in robotics from the University of Tokyo in 2016. Stéphane has done research on humanoid locomotion at CNRS and was locomotion team lead at ANYbotics AG before joining Inria, where he currently investigates applications of machine learning to robotics, particularly at the interface of computer vision and motion control. Stéphane advocates for open-source robotics and contributes to projects including Upkie wheeled bipeds, robot_descriptions.py, and qpbenchmark.

This talk presents recent algorithmic advances that accelerate robot dynamics simulation and control by exploiting structure and reducing computational complexity. We introduce low-order algorithms for simulating constrained rigid-body systems, including novel recursive methods that can address kinematic loops and are robust to singularities. Building on these foundations, we show how these algorithms enable fast and scalable solutions to core robotics problems such as differential inverse kinematics, frictional contact simulation, and trajectory optimization. Together, these contributions enable efficient online computation for control and planning in high-dimensional robotic systems.

Ajay Sathya received his Bachelor's degree from NITK Surathkal, India, in 2016, and both his Master's and Ph.D. degrees from KU Leuven, Belgium, in 2018 and 2023 respectively, all in mechanical engineering. He is a Marie Skłodowska-Curie Postdoctoral Fellow in the Willow research team at Inria and École Normale Supérieure, Paris. His current research focuses on developing scalable algorithms for differentiable simulation and control of high-dimensional robots operating in contact-rich environments. His work has been recognized with awards including the IEEE RA-L 2024 Best Paper Award and IROS 2023 Best Paper Finalist. He was also selected as a Robotics: Science and Systems Pioneer in 2024.

The formulation of robotics tasks as optimization problems is ubiquitous, spanning from traditional sense-act architectures to contemporary policy learning paradigms. A significant challenge, however, is the non-convex nature of these problems, which often leads to suboptimal solutions when using standard local solvers without adequate initialization. Consequently, current robotics pipelines often employ sophisticated initialization schemes, large-scale computational searches, or mitigation strategies to address suboptimal solutions. This talk will discuss a more principled alternative for achieving global optimality, rooted in polynomial optimization, specifically sums-of-squares and moment hierarchy methods. I will present recent contributions aimed at lowering the barrier to entry for these methods by automating the formulation and tightening of their relaxations, enhancing scalability, moving beyond poorly conditioned monomial feature functions, and developing structure-exploiting solvers for the resulting semidefinite programs.

Frederike Dümbgen is an incoming assistant professor at Carnegie Mellon University in Mechanical Engineering. She is currently a researcher in the Willow team of Inria Paris, where she has been working on optimization for robotics since May 2024 under the Marie Curie Postdoctoral Fellowship LiftMeUp. Before joining Inria, she spent two years as a postdoctoral fellow at the Robotics Institute of the University of Toronto, collaborating with Prof. Timothy D. Barfoot on certifiable optimization. Frederike holds a Ph.D. in Computer and Communication Sciences from École Polytechnique Fédérale de Lausanne (EPFL), Switzerland. She was recognized as an RSS Pioneer in 2024, as a Google Women Techmaker in 2020, and has served as associate co-chair of the RAS technical committee on model-based optimization since 2024.