This talk explores how partial differential equation (PDE)–based physics priors can provide a foundation for scalable and generalizable algorithms in robot motion learning. Rather than searching over discrete graphs or samples, it formulates and learns the solution to the motion-planning problem as a continuous value function governed by Hamilton–Jacobi (HJ) PDEs. These methods enable self-supervised value-function learning without reliance on expert trajectories or trial-and-error interaction. The learned value functions yield fast inference of motion plans and demonstrate strong scalability across complex, high-dimensional, and constraint-rich navigation and manipulation tasks. The talk also introduces an HJ PDE–derived mapping representation that unifies perception and planning: unlike occupancy grids or signed distance fields, it encodes motion-feasible geometry in a form naturally structured for continuous decision-making. Together, these developments outline a unified, numerically grounded framework for robot motion planning and control through the lens of physics-informed learning.

Robots show great promise if they can get out of the lab into the field and go beyond a single-operator per robot paradigm. However, the unstructured nature of the real-world requires nuanced decision making of the robot. In this talk I will outline some of our approaches, progress, and results on multi-modal sensing, providing nuanced perception inputs, as well as navigation in difficult terrain, and future directions of our research.

Deploying mobile robots in unstructured outdoor environments remains a fundamental challenge, requiring the ability to robustly perceive complex terrains, pedestrian flows, and general traffic rules. To effectively serve humans, especially older adults, these robots must go beyond simple navigation to also understand human behavior and enhance personal mobility. In this talk, I will review our previous approaches for long-range outdoor navigation, with a focus on scene understanding and planning. Then, I will present a high-level overview of what we are currently working on, where I aim to apply these navigation technologies to develop companion robots that support older adults.

Humanoid robots are entering human-centric environments, where they must not only move well but also understand people and interact safely through physical contact. In this talk, I will present two complementary directions toward human-centered embodied intelligence. First, I will introduce GentleHumanoid, a whole-body control policy that combines motion tracking with compliant, tunable force regulation, enabling contact-rich behaviors such as gentle hugging, assistive support, and safe object interaction on the Unitree G1. Second, I will show how large language models can be grounded in 3D human motion for behavior understanding and planning, highlighting ChatPose and ChatHuman as steps toward systems that interpret actions, anticipate intent, and connect high-level reasoning to executable motion. I will close with future directions on scaling human–humanoid interaction data, developing vision-language-action models for long-horizon interaction, and incorporating muscle-driven modeling for more realistic and adaptive humanoids. 

Despite decades of advancement, autonomous driving systems have not met the high expectations set by many. What’s missing is physical intelligence - the ability of AI systems to reason, react, and adapt in real time, while operating safely and effectively within the laws of physics. In this talk, I will first examine which hurdles have turned out to be more formidable than expected, and share our research on how to refine testing methodologies to advance the safety of autonomous vehicles. I will then show how high-speed autonomous racing provides a unique proving ground to test the boundaries of AI’s physical capabilities. Leveraging more than a decade of experience in high-speed autonomous racing, particularly with the full-scale Cavalier Autonomous Racing Indy car and the F1tenth platform, I will demonstrate how racing at high speeds and in close proximity to other vehicles exposes unsolved challenges in perception, planning, and control. I will recount our journey from the lab to lap times, and the rigorous engineering required to build a full-scale autonomous racecar from scratch. Despite progress, autonomous racing has yet to match the skill of expert human drivers or master the complexity of dense, multi-car competition; indicating that we still have several more laps to go on our path toward artificial general “driving” intelligence.

This talk explores the interplay between model-based guarantees and learning-based flexibility in the control of dynamical systems. I begin with safety-critical control using control barrier functions (CBFs), highlighting that while CBFs enforce state constraints, they may induce unstable internal dynamics that renders the system 'unsafe'! To address this, I introduce conditions under which CBF-based safety filters also ensure boundedness of the full system state. I then transition to learning representations of hybrid dynamical systems. I present a framework that learns continuous neural representations by exploiting the geometric structure induced by guards and resets, enabling accurate flow prediction of hybrid systems without explicit mode switching. Finally, I discuss generative learning approaches for control. Through applications to legged robotics, I illustrate how a generative sensorimotor model can generalize beyond the training distribution of pure locomotion and manipulation to achieve whole-body control. Together, these results highlight how structure, geometry, and learning can bridge safety guarantees and expressive control for complex dynamical systems.

How can we build robots that meaningfully assist people with mobility limitations in their daily lives? To support complex caregiving tasks such as robot-assisted feeding, transferring, bathing, and meal preparation, robots must physically interact with people and objects in dynamic, unstructured environments. In this talk, I will present an overview of various projects from my lab that showcase fundamental advances in the field of physical robotic caregiving. I will highlight how we design stakeholder-informed systems, from simulation platforms to real-world deployments, by integrating multimodal perception, user feedback, and adaptive algorithms. Together, these efforts move us closer to creating caregiving robots that are not only technically capable, but are also responsive to the real needs of people in care settings.

In this talk, I am going to share our recent work 𝚿0, an open foundation model to address challenging humanoid loco-manipulation tasks. While existing approaches often attempt to address this fundamental problem by co-training on large and diverse human and humanoid data, we argue that this strategy is suboptimal due to the fundamental kinematic and motion disparities between humans and humanoid robots. Therefore, data efficiency and model performance remain unsatisfactory despite the considerable data volume. To address this challenge, our method decouples the learning process to maximize the utility of heterogeneous data sources. Specifically, we propose a staged training paradigm with different learning objectives: First, we autoregressively pre-train a VLM backbone on large-scale egocentric human videos to acquire generalizable visual-action representations. Then, we post-train a flow-based action expert on high-quality humanoid robot data to learn precise robot joint control. Our research further identifies a critical yet often overlooked data recipe: in contrast to approaches that scale with noisy Internet clips or heterogeneous cross-embodiment robot datasets, we demonstrate that pre-training on high-quality egocentric human manipulation data followed by post-training on domain-specific real-world humanoid trajectories yields superior performance. Extensive real-world experiments demonstrate that 𝚿0 achieves the best performance using only about 800 hours of human video data and 30 hours of real-world robot data, outperforming baselines pre-trained on more than 10 times as much data by over 40% in overall success rate across multiple tasks. We will open-source the entire ecosystem to the community, including a data processing and training pipeline, a humanoid foundation model, and a real-time action inference engine.

Moravec’s Paradox observes that AI systems have struggled far more with learning physical actions than symbolic reasoning. Yet just recently, there has been a tremendous increase in the capability of AI-driven robotic systems, reminiscent of the early improvements in language modeling capabilities a few years ago. In this talk, we provide mathematical evidence that learning in continuous-control settings, like robotics, can be exponentially more challenging than in discrete settings, like language, unless certain key algorithmic design choices are made - effectively, mathematical evidence for Moravec’s claim. We then show that two of the key innovations in modern robot learning - action chunking, and the use of generative models, such as diffusion models, to parametrize robot actions - can be interpreted as directly mitigating the mechanisms underlying this difficulty. Our perspective runs contrary to many popular justifications for the two methods, such as capturing multi-modality present in mixed-quality training data. Finally, if time permits, we will describe a new family of interventions, at the level of deep learning optimization, that provide yet another lever for addressing the same challenges.

One of the biggest questions in robotics today is how to simply and easily teach robots new tasks. In this talk, we explore whether we can teach robots the same way we teach humans: by having the robots watch humans execute the tasks. We show that despite the huge sensing and morphology gap between robots and humans, robots can understand skill sequencing, object detection, and task geometry just from watching human demonstrations.