Welcome to my website!
Leonel Rozo
Head of the Physical Holistic Intelligence Lab Italian Institute of AI for Industry (AI4I) |
I love machine learning and robotics!
When I started my studies in Colombia, I got excited about how the future in AI and robotics would look like. Passion, endurance and great mentors have led me to work on what I love: research and lead projects at the intersection of machine learning, differential geometry, control theory, and robotics. I envision a world where smart robots augment our capabilities at physical, cognitive, and social levels! AI and robots that can easily assist, help, and empower people in a large diversity of scenarios and tasks. To make this dream come true, I spend my days working on machine learning tools that allow robots to safely and smartly interact with their environment and humans. Day by day, I hit the books and investigate machine learning methods, robot control, differential geometry and optimization. |
News and updates!
ai4i.it[October, 2025] Recently moved to Turin (Italy) to start a new adventure as the director of the Physical Holistic Intelligence (PHI) lab at AI4I, many news coming soon!
[July, 2025] New paper accepted at IEEE T-Ro! We have a quite fast and robust formulation of flow matching for learning visuomotor robot policies! What could be cooler than that? Well, we made it Riemannian as well so that we can learn policy distributions on non-Euclidean spaces, necessary when modeling end-effector motions or variable impedance policies. Check out our paper "Fast and Robust Visuomotor Riemannian Flow Matching Policy". Great job, Haoran!
[June, 2025] Flow matching hits again! Haoran just got his paper on "Towards Safe Imitation Learning via Potential Field-Guided Flow Matching" accepted at IROS'25! It turns out that we can learn safety constraints from demonstrations and design a potential field to guide a flow matching model during inference, so that a robot skill is safely execute! Don't miss the details and check his paper!
[May, 2025] So happy for Andrea's first paper at ICML'25! Are you familiar with the geometry structure arising from dynamical systems, e.g., the Symplectic manifold of Hamiltonian systems? Well, Andrea leveraged a more general geometric structure, the contact manifold, to model and control dynamical systems, naturally including non-conservative dynamics. Check out our paper "Geometric Contact Flows: Contactomorphisms for Dynamics and Control".
[May, 2025] Hadi just put the cherry on top his PhD thesis cake! Check out our new IJRR paper on "Extended Neural Contractive Dynamical Systems: On Multiple Tasks and Riemannian Safety Regions"! We bring very cool extensions for our ICLR'24 paper! NCDS can now handle multiple tasks via conditioning, it carries out latent space obstacle avoidance via pullback metrics, and we provide a thorough analysis on regularization schemes for learning contractive dynamical systems.
[April, 2025] We are so back at R:SS! Ken-Joel, our MSc. student, got his first paper accepted! We design a safe and robust obstacle avoidance approach for Neural Contractive Dynamical Systems (NCDS) that preserves contraction guarantees under diffeomorphisms! Check out our paper "Diffeomorphic Obstacle Avoidance for Contractive Dynamical Systems via Implicit Representations".
[January, 2025] What a way to start the year! Our first paper at AISTATS! This time, we could not be more Riemannian! Check out our new paper "Riemann2: Learning Riemannian Submanifolds from Riemannian Data".
[May 2, 2024] So happy we made it in ICML'24! If you are curious about how motion taxonomies can be leveraged to build taxonomy-aware embeddings in hyperbolic manifolds, then check out our new paper on "Bringing Motion Taxonomies to Continuous Domains via GPLVM on Hyperbolic manifolds".
[January, 2024] Good news from ICRA'24! My partner in crime and I will be talking about "The Single Tangent Space Fallacy", a paper about technically-sound practices when applying Riemannian geometry in robot learning problems. See you in Japan!
[January, 2024] Super happy about Hadi's paper accepted as a spotlight in ICLR'24! We propose "Neural Contractive Dynamical Systems (NCDS)", a new method to learn low-dimensional dyn. systems in a VAE latent space with contraction-stability guarantees, and it works in the SE(3) Lie group!
[January, 2024] ICRA'24 will host the second edition of our tutorial on Riemannian manifolds applied in robot learning, optimization a control! We are so excited to meet you in Japan and talk about what is and how to exploit Riemannian methods in robotics problems.
[July, 2025] New paper accepted at IEEE T-Ro! We have a quite fast and robust formulation of flow matching for learning visuomotor robot policies! What could be cooler than that? Well, we made it Riemannian as well so that we can learn policy distributions on non-Euclidean spaces, necessary when modeling end-effector motions or variable impedance policies. Check out our paper "Fast and Robust Visuomotor Riemannian Flow Matching Policy". Great job, Haoran!
[June, 2025] Flow matching hits again! Haoran just got his paper on "Towards Safe Imitation Learning via Potential Field-Guided Flow Matching" accepted at IROS'25! It turns out that we can learn safety constraints from demonstrations and design a potential field to guide a flow matching model during inference, so that a robot skill is safely execute! Don't miss the details and check his paper!
[May, 2025] So happy for Andrea's first paper at ICML'25! Are you familiar with the geometry structure arising from dynamical systems, e.g., the Symplectic manifold of Hamiltonian systems? Well, Andrea leveraged a more general geometric structure, the contact manifold, to model and control dynamical systems, naturally including non-conservative dynamics. Check out our paper "Geometric Contact Flows: Contactomorphisms for Dynamics and Control".
[May, 2025] Hadi just put the cherry on top his PhD thesis cake! Check out our new IJRR paper on "Extended Neural Contractive Dynamical Systems: On Multiple Tasks and Riemannian Safety Regions"! We bring very cool extensions for our ICLR'24 paper! NCDS can now handle multiple tasks via conditioning, it carries out latent space obstacle avoidance via pullback metrics, and we provide a thorough analysis on regularization schemes for learning contractive dynamical systems.
[April, 2025] We are so back at R:SS! Ken-Joel, our MSc. student, got his first paper accepted! We design a safe and robust obstacle avoidance approach for Neural Contractive Dynamical Systems (NCDS) that preserves contraction guarantees under diffeomorphisms! Check out our paper "Diffeomorphic Obstacle Avoidance for Contractive Dynamical Systems via Implicit Representations".
[January, 2025] What a way to start the year! Our first paper at AISTATS! This time, we could not be more Riemannian! Check out our new paper "Riemann2: Learning Riemannian Submanifolds from Riemannian Data".
[May 2, 2024] So happy we made it in ICML'24! If you are curious about how motion taxonomies can be leveraged to build taxonomy-aware embeddings in hyperbolic manifolds, then check out our new paper on "Bringing Motion Taxonomies to Continuous Domains via GPLVM on Hyperbolic manifolds".
[January, 2024] Good news from ICRA'24! My partner in crime and I will be talking about "The Single Tangent Space Fallacy", a paper about technically-sound practices when applying Riemannian geometry in robot learning problems. See you in Japan!
[January, 2024] Super happy about Hadi's paper accepted as a spotlight in ICLR'24! We propose "Neural Contractive Dynamical Systems (NCDS)", a new method to learn low-dimensional dyn. systems in a VAE latent space with contraction-stability guarantees, and it works in the SE(3) Lie group!
[January, 2024] ICRA'24 will host the second edition of our tutorial on Riemannian manifolds applied in robot learning, optimization a control! We are so excited to meet you in Japan and talk about what is and how to exploit Riemannian methods in robotics problems.
