Program

Abstract

Equipping AI and robotic systems with the ability to understand human behavior is essential for enabling them to assist people across a wide range of everyday applications. This need is more pressing than ever: the heaviest consumers of such knowledge are no longer perception systems alone, but robot policies that must learn to act in the physical world. Yet the high-quality 3D human motion data required to learn this knowledge remains extremely scarce.

In this talk, I will present our lab's efforts to scale and enrich 3D human motion data by capturing everyday movements and natural human-object interactions, with the ultimate goal of teaching robots to move like humans.

I will first introduce our multi-year effort in building multi-camera capture systems, from Panoptic Studio to ParaHome, and most recently OmniRoboHome, a new system designed to capture human-object interactions in natural home environments. Next, I will present a complementary direction: learning everyday interactions and affordances from generative image and video models, which offer a scalable source of human behavior priors that capture systems alone cannot reach. Finally, I will discuss the missing pieces that vision and image models cannot provide, including physics, contact, and the gap across diverse embodiments, together with our recent efforts to fill them.

Bio

Hanbyul Joo is an associate professor at Seoul National University (SNU) in the Department of Computer Science and Engineering. Before joining SNU, Hanbyul was a Research Scientist at Facebook AI Research (FAIR), Menlo Park. Hanbyul received his PhD from the Robotics Institute at Carnegie Mellon University. Hanbyul is a recipient of the Best Student Paper Award at CVPR 2018.

Abstract

Humanoid robots offer a versatile platform for performing a wide range of tasks in complex, human-centered environments. Their human-like embodiment makes human data a promising foundation for scalable skill learning. However, translating human skills into robot control requires bridging cross-embodiment gaps, particularly in dynamics and actuation.

In this talk, we present a general recipe for learning humanoid skills from human data with physics grounding. Human data provides high-level intent and skill structure, while a physics-grounded layer translates them into motor-level actions. This approach enables versatile, adaptive, and robust skill learning across a range of loco-manipulation and dexterous manipulation tasks.

Abstract

For mobile robots to work full shifts, they need to cope with the messiness and unpredictability of the real world for hours at a time without accumulating failures—potentially even in open-world settings. Robust perception and 3D/4D scene understanding therefore remain key challenges. In this talk, I will present several recent works on real-time perception that exploit complementary sensing modalities, including event cameras, to improve robustness. I will also discuss approaches that lift open-vocabulary vision-language features into 3D representations, enabling semantic navigation and higher-level task execution on drones and other mobile robots.

Bio

Stefan Leutenegger is Associate Professor of Mobile Robotics at ETH Zurich, where he leads the Mobile Robotics Lab. Before joining ETH in 2025, he held positions at the Technical University of Munich and Imperial College London. He received his BSc, MSc, and PhD degrees from ETH Zurich. His research focuses on robust perception and autonomy for mobile robots, including visual-inertial navigation, SLAM, 3D scene understanding, and robot learning.

Abstract

The ability to build an actionable understanding of the environment of a robot is crucial for autonomy and prerequisite to a large variety of applications, ranging from home, service, care and consumer robots to autonomous vehicles, augmented reality, and disaster response. Notably, much of this depends on long-term operation in human-centric domains that are complex, widely diverse, and highly dynamic. This talk presents an autonomy pipeline to address these challenges. At the core, I argue that time, and thus memory, dynamics, and adaptation, should be an integral component of AI systems. I will introduce methods to capture the present and past of complex and dynamic scenes through symbolic abstractions, which further facilitate predicting future scene outcomes. I will show how robots as embodied agents can leverage our actionable scene representations and predictions to complete tasks such as actively gathering data that helps them improve their scene models and perception capabilities, and how all these tools can combine for robots to fully autonomously learn over time.

The presented methods are demonstrated on-board fully autonomous aerial and ground robots, run in real-time on the limited hardware available, and are released as open-source software.

Bio

Lukas Rosenberger Schmid (formerly Schmid) is a Tenure-Track Professor of Machine Intelligence at UTN. Before that, he was a Research Scientist and Postdoctoral Fellow at the SPARK Lab led by Prof. Luca Carlone at MIT, and a Postdoctoral Researcher at the Autonomous Systems Lab (ASL) led by Prof. Roland Siegwart at ETH Zürich. He earned his PhD in 2022 from ASL at ETHZ, where he also was a visiting researcher at the Microsoft Spatial AI Lab led by Prof. Marc Pollefeys. His work has been recognized by several honors, including RSS Pioneers 2025, NOKOV New Generation Star at IROS 2025, the RSS Outstanding Systems Paper Award 2024, two ETH Medals for outstanding PhD and M.Sc. Theses, the Willi Studer Prize for the best graduate of the year at ETHZ, the first place in the 2024 Hilti SLAM challenge, and a Swiss National Science Foundation (SNSF) Postdoc Fellowship.

His research focuses on embodied spatio-temporal AI for human-centric robot autonomy. This includes research on scene representations and abstraction, on detection, prediction, and understanding of moving and changing entities, active perception and information gathering, as well as lifelong learning for continuous adaptation to the robot environment, embodiment, task, and human preference.

Abstract

Trajectory optimization can make robots reach, dodge, and walk, but on high-dimensional systems such as humanoids it typically takes seconds to minutes and requires a good initial guess. This talk presents FLINT (Finite-time heat-fLow INTegration), a heat-flow method that needs no initial guess: it can start from any curve that connects the start to the goal. Drawing on the theory of fixed-time stability, FLINT discretizes the heat flow with a step size that grows as the flow approaches a solution, and we prove that it converges in a number of steps that is bounded independently of the requested accuracy. Across arms and humanoids, FLINT is 30 to 530 times faster than state-of-the-art direct, DDP-based, and earlier heat-flow solvers, with equal or better success when every solver is judged by the same criterion. FLINT plans collision-free arm motions among obstacles given as sensed point clouds in milliseconds, solves a 22-joint humanoid in 14 ms, and produces seven walking gaits for the Talos humanoid, including stair climbing, each from a standing start in under 0.3 s.

Bio

Ram Vasudevan is a professor in Mechanical Engineering and Robotics at the University of Michigan. He received a BS in Electrical Engineering and Computer Sciences, an MS degree in Electrical Engineering, and a PhD in Electrical Engineering all from the University of California, Berkeley. He is a recipient of the NSF CAREER Award, the ONR Young Investigator Award, and the 1938E Award from the University of Michigan. His work has received best paper awards at the IEEE Conference on Robotics and Automation, the ASME Dynamics Systems and Controls Conference, and IEEE International Conference on Biomedical Robotics and Biomechatronics, and has been finalist for best paper at Robotics: Science and Systems.

Abstract

Agentic systems have been widely deployed in the digital space and are now entering the physical world. In this talk, I will discuss our recent efforts in bringing such systems into robotic manipulation, with their strengths and limitations. I will in particular focus on how coding agents may connect to robotic task planning, and present our thinking on the connections between learning systems and structured representations across multiple levels of abstraction.

Bio

Jiajun Wu is an Assistant Professor of Computer Science and, by courtesy, of Psychology at Stanford University, working on computer vision, machine learning, robotics, and computational cognitive science. Before joining Stanford, he was a Visiting Faculty Researcher at Google Research. He received his PhD in Electrical Engineering and Computer Science from the Massachusetts Institute of Technology. Wu's research has been recognized through the IJCAI Computers and Thought Award, the Young Investigator Programs (YIP) by ONR and by AFOSR, the Early Career Program (ECP) by ARO, the NSF CAREER award, the Okawa research grant, and paper awards and finalists at ICCV, CVPR, SIGGRAPH Asia, ICRA, CoRL, and IROS.

Abstract

Agentic reasoning for robots is rapidly becoming a reality, allowing flexible natural language interaction with human operators and enabling a wide range of navigation, object handling and recall tasks in a variety of settings. In this talk, Prof. Waslander will discuss the ongoing efforts in his lab to make useful agentic robots for the warehouse automation, by integrating open world perception with agentic reasoning for reliable open world navigation, and by adding multi-faceted memory - spatial, descriptive and visual - to enable long-term experience recall for visual question answering. Together, these advances enable wide variety of spatial, semantic, functional and temporal tasks.

Bio

Prof. Steven Waslander is a leading authority on autonomous robotics, including self-driving cars and multirotor drones. He received his B.Sc.E.in 1998 from Queen’s University, his M.S. in 2002 and his Ph.D. in 2007, both from Stanford University in Aeronautics and Astronautics. He was recruited to the University of Waterloo from Stanford in 2008, where he led the Autonomoose project, the first self-driving car to be tested on public roads by a Canadian university. In 2018, he joined the University of Toronto Institute for Aerospace Studies (UTIAS), and founded the Toronto Robotics and Artificial Intelligence Laboratory (TRAILab).

Abstract

Robotic systems are increasingly expected to work closely and smoothly around humans, but the dynamic nature of human behavior makes this hard: a robot must continuously predict and respond to evolving human behavior rather than merely react to the current observation. This talk studies this challenge through the Robot Follow-Ahead (RFA) problem, where a robot must move in front of a walking human to maintain visibility of high-priority frontal areas such as facial expressions. I'll present a progression of methods that couple human motion prediction with robot path planning in a closed loop, with the help of diffusion models and Bayesian methods.

Bio

Qingyuan Jiang is a Machine Learning Engineer at Apple Camera Incubation team. He received his PhD degree from the University of Minnesota, advised by Prof. Volkan Isler. His research lies at the intersection of human motion prediction and robot motion planning, with a focus on perception-oriented human-robot interaction. His PhD work was supported in part by the 3M Fellowship, and an NSF NRI Award.

Bio

Lloyd Lei is a CMU MSCS student who is currently on leave to build his startup, ArenaLabs, as its founder and CEO. The company just finished a $750K angel round and is building the largest across-embodiment dataset in the world.

He was one of the youngest YC China partners. He has held research internships at ByteDance (TikTok), Broadcom, and the Robotics and AI Institute (RAI), and participated in StarBot’s seed round. Before his academic career in AI, he was a visiting researcher at Lawrence Berkeley National Laboratory, studying particle physics.

His current research interests and investments mainly focus on Whole-Body Control, real-time and low-latency inference, and across-embodiment representation learning.