A longstanding goal in robotics is to build agents that learn from the world and assist people in everyday tasks across homes, factories, and streets. This talk outlines a path to open world autonomy that learns continuously, reasons with language and vision, and closes the loop from perception to action. I will present representations that capture objects, relations, and articulation, online learning that adapts during deployment without forgetting, and uncertainty-aware decision making that knows when to ask for clarification, seek information, or recover. I will also discuss data and model efficiency in policy learning for long-horizon tasks, including from demonstrations, teleoperation, and world models for rapid offline adaptation. I will conclude with a discussion of safety, fairness, and responsible deployment, so that learning-enabled autonomy earns trust and delivers value to society.

Foundation models and large-scale learning are transforming robotics: robots can now be programmed through language, vision, and demonstrations rather than hand-crafted code. Yet this new paradigm raises a fundamental question: how do we guarantee safety in open, dynamic environments? In this talk, I will present recent work on combining semantic understanding and generative policies with safety-critical control frameworks. Examples include safe crowd navigation using diffusion models with human motion prediction, semantic safety filtering for manipulation, and swarm control through natural-language instructions with built-in safeguards. These results highlight both the new opportunities and challenges of deploying internet-scale AI in robotics. I will conclude with a roadmap for trustworthy, scalable, and safe autonomy in human-centered environments.

Soft robots, inspired by the biological systems, face the issue of being prone to damage. However, biological entities have self-repair capabilities—a feature we’ve introduced in soft robots to foster renewed confidence in their reliability. Our technological advancements enable these robots to self-heal, enhancing their durability and extending their operational lifespan. This innovation not only increases reuse but also allows for recycling and is based on bio-sources, contributing to sustainability. We’ve revolutionized the entire value chain by developing materials that surpass mere coatings; they form structural 3D components with wide range of mechanical, conductive, and magnetic properties. These materials are compatible with extrusion and molding techniques as well as multi-material printing—processes typically unsuitable for traditional network polymers and delamination risks at material interfaces. Our breakthroughs innovations include self-repairing robotic grippers with integrated sensors that not only detect but also respond to damage. Currently we are maturing the technology in order to realize a deeptech spinoff Valence Technologies, commercialising self-closing suction cups and self healing bike and car tires.

This talk will discuss recent work aimed at advancing the autonomy of marine robots operating in complex environments. First, to achieve the situational awareness needed for autonomous inspection and precise physical intervention, I will discuss research that aims to produce accurate, high-definition 3D maps of underwater structures using wide-aperture multi-beam imaging sonar. Second, I will discuss research intended to help marine robots make safe and efficient navigation decisions under both epistemic and aleatoric uncertainty. To address the former, sonar-equipped underwater robots use "virtual maps" as a tool to support accurate map-building under localization uncertainty. To address the latter, we employ distributional reinforcement learning to help lidar-equipped unmanned surface vehicles navigate congested and disturbance-filled environments. Our results include several open-source algorithm implementations and benchmarking tools.

Robotics is moving steadily toward greater reliance on AI, massive datasets, and ever-increasing computation, often with the implicit assumption that more power and more complexity will yield more intelligence. While these approaches have delivered impressive capabilities, they also come at the cost of energy, scalability, and accessibility. In contrast, biological intelligence is strikingly frugal, achieving robust sensory–motor coordination and adaptive behavior under severe energy and computational constraints. One key mechanism behind this frugality is embodied intelligence: musculoskeletal structures, compliant materials, and morphological design allow the body itself to offload part of the computation mechanically, reducing the burden on centralized control. This principle of morphological computation shows that perception and action emerge not only from neural processing, but from the tight coupling of body, environment, and control. Looking ahead, we can envision the extreme case of electronics-free soft robots—robots whose behavior is programmed mechanically, not digitally.

Reviewing the progress in artificial intelligence over the past decade, AI systems have achieved human-level understanding and reasoning at the semantic level and have been widely adopted in Internet scenarios. Nevertheless, when it comes to understanding and interacting with the physical world, AI systems still face significant challenges. This reveals an important issue: relying solely on the semantic-level concepts learned by large models is far from sufficient for embodied applications -- current AI systems lack an effective way to comprehend the physical world. This research considers the processes of object manufacturing and manipulating from a cognitive perspective and introduces the idea of “digital gene”  representing the commonalities and differences of similar objects through programmed and parameterized methods, providing a computable, unambiguous, and highly generalizable structured abstraction at the level of physical concepts. In this manner, it offers machine intelligence a portal to perceive, reason about, and interact with the physical world.

Event cameras are bio-inspired vision sensors with much lower latency, higher dynamic range, and much lower power consumption than standard cameras. This talk will present current trends and opportunities with event cameras, ranging from robotics to virtual reality and smartphones, as well as open challenges and the road ahead.

"Robots are usually designed for utility — to execute tasks efficiently, reliably, and with precision. Yet when freed from function, robots can take on entirely new roles: as cultural artifacts, performers, and even works of art. What happens when a machine is no longer judged only by how well it works, but by how deeply it can move us, provoke questions, and spark imagination? In this keynote, Dennis Hong shares a journey of reimagining robotics through unexpected contexts — from COSMO, a robot featured in a Hollywood blockbuster film, to experimental installations where play, presence, and emotion take center stage. These experiences reveal how the artistic lens can drive new forms of engineering creativity and expand our understanding of what robots can be. Looking ahead, Hong offers a first glimpse of Aequor Triformis, a new robotic artwork inspired by natural fluidity and designed to blur the line between mechanism and organism. Together, these explorations invite us to view robots not only as tools of science and industry, but as mirrors of human imagination — expanding the future of robotics from work, to wonder."

Autonomous motor skill learning and control are central challenges in the development of intelligent robotic systems. Imitation learning offers an efficient approach, enabling robots to acquire new skills from human guidance while reducing the time and cost of manual programming. Traditional approaches to robot learning from demonstration tend to render the robot a passive learner, confined largely to motion planning derived from the current observations. To progress beyond the traditional paradigm of imitation learning, it is essential to develop methods that allow robots to continuously learn new skills and to refine previously learned ones, if necessary, particularly in uncertain or dynamic environments. This may require the ability to reason about the robot’s own actions, or to extend its knowledge through proactive and interactive engagement with humans.

Recent advent of artificial intelligence and rapid progress on motion capacity of humanoid robots are bringing a strong attention expecting industrial and social applications. Despite their highly dynamic motion ability, humanoid robots need further improvements in motions involving contacts with all over their body. We humans generate such contact-rich motions with ease while solving the complex problem of combined discrete contact sequence and continuous dynamic motions is extremely hard. In this talk, I will present ongoing research to address the following challenges: collecting whole-body contact motion data from humans, learning behaviors from those data and synthesizing multi-contact humanoid motions.

Deploying Embodied AI and multi-agent systems is critically hampered by challenges in perception uncertainty and robust decision-making in various scenarios. This talk presents novel uncertainty quantification and robust multi-agent reinforcement learning (MARL) frameworks that directly confront this challenge. First, we introduce an uncertainty quantification method for deep learning-based perception and prediction models. Building upon this, we provide a theoretical analysis of MARL under state uncertainties, leading to a provably robust algorithm that can withstand worst-case uncertainties. Furthermore, we leverage control theory with robust MARL framework to achieve a critical balance between provable safety and high operational efficiency. We demonstrate the power of this research in the context of connected autonomous vehicles, validating our framework in both high-fidelity simulators and on real-world hardware testbeds. The talk will conclude by outlining a forward-looking research agenda for creating the next generation of trustworthy, cooperative AI systems.

The Vision-Language-Action (VLA) paradigm has significantly advanced robotic control through Internet-scale pre-training. However, its application to real-world manipulation tasks, particularly those requiring high precision in contact-rich scenarios or dealing with complex dynamics, is often limited by a lack of fine-grained physical grounding. To address this, we propose a Knowledge-Guided Tactile VLA framework that enhances traditional vision-language-action models with robust physical reasoning capabilities through tactile sensing and world modeling. Our Unified Digital Physics System (UDPS) incorporates tactile perception with physical knowledge prior via a novel tokenization scheme that encodes geometry, physics, and tactile cues into a unified representation. The cross-domain alignment distilled from geometry invariances substantially improving sim-to-real transfer for contact-rich manipulation. Simultaneously, physical token enables the modelling of dynamic and complex physical process, including soft-body deformation and contact transitions. The framework is rigorously validated in two demanding tasks: precision 3C assembly and humanoid handkerchief dancing. In 3C assembly, UDPS taking tactile feedback as position offset in sim-to-real transfer and achieves sub-millimeter precision in connector mating in a zero-shot manner. For handkerchief manipulation, the physical tokens models complex fabric dynamics, enabling stable rhythmic motions through whole-body coordination. These results demonstrate the critical importance of integrating physical knowledge and tactile sensing for solving complex, contact-rich manipulation tasks in real-world environments without real-world fine-tuning.

The employment of robotic platforms and autonomy might significantly increase the efficiency and reduce the risk to humans in search and rescue missions. Motivated by insights from the autonomous collective foraging behaviors performed by animals in aquatic environments, this talk introduces models and provable strategies from control theory and robotics towards bio-inspired autonomous search operations. The bio-inspired methods generalize to a Bayesian learning framework where insights from biology are well justified by systems theory such as reachability, consistency, and optimality.  Experimental effort with promising results demonstrates that bio-inspired autonomy might be preferred in aquatic environment that features severe limitation in communication, localization, and power consumption

Soft robotics has made remarkable advances in developing deformable functional materials for locomotion, manipulation, and other forms of morphological adaptation such as self-morphing, self-healing, and mechanical growth. While these technologies have opened up new applications for robotics, they also present novel challenges in sensing, modelling, planning, and control. Due to the inherent complexity of systems based on flexible and continuum mechanics — and the wide range of interactions with their environments — conventional methods often fall short, making novel approaches rooted in advanced machine learning essential. In this talk, I will introduce several projects in our laboratory that leverage sensorized soft robots and machine learning to tackle these challenges. I will also present the concept of “Super Embodied Intelligence” as a new research framework for realizing the next generation of intelligent robots and its technological underpinnings. As research in soft robotics and functional materials progresses, we are witnessing a fusion of the informational and physical entities. Within this context, where new forms of embodied intelligence are emerging, I will discuss how rapidly evolving fields such as machine learning can accelerate this development. Moving beyond traditional notions of bodily control and AI as purely computational, this approach explores the potential for new forms of intelligence in which the body itself becomes an active site for information processing and generation.

Wearable and surgical robots have the potential to transform human health and performance, but their development has been slowed by two persistent challenges: they are often bulky and confined to lab settings, and they lack autonomy to effectively collaborate with humans. Our work on high-torque density motors enables compact, lightweight exoskeletons and neurosurgical robots. On the control side, we introduce a physics-informed learning-in-simulation framework, combined with deep reinforcement learning, that creates adaptive controllers without costly human experiments. This approach, published in Nature, allows robots to understand human intention and act with greater autonomy. Together, these advances move robotics beyond lab prototypes toward real-world systems that make movement easier, surgery safer, and healthcare robotics more accessible.

In the pursuit of scalable, socially aware, and safety-critical autonomous systems, our recent research has focused on integrating learning, planning, and control across aerial, ground, and maritime robotic platforms. Central to this effort is the fusion of model-based and data-driven approaches, enabling robust decision-making in dynamic and uncertain environments, seamless multi-robot coordination, and the ability to learn from human demonstrations. This talk will highlight recent advances in three key areas: 1) interaction-aware navigation among other robots and humans, using sampling-based model predictive control, socially compliant behavior learning, and semantic mapping; 2) real-time task and motion planning for teams of mobile manipulators through expert demonstrations, physically grounded plans, and whole-body control; and 3) decentralized 6-DoF manipulation of cable-suspended loads by a team of drones using multi-agent reinforcement learning. These contributions advance the frontier of scalable autonomy in dynamic, multi-agent environments across diverse robotic platforms.

Soft growing robots, often referred to as vine robots, represent a new class of continuum robots that achieve locomotion by extending their body through tip eversion, much like the growth of a plant vine. This simple yet powerful principle enables robots to navigate confined and cluttered environments without causing significant disturbance to their surroundings. Despite their promise, early implementations have faced key challenges that limit their deployment in real-world scenarios, including restricted steering, difficulty in mounting sensors and tools at the tip, challenges in controlled retraction, and robustness under diverse operating conditions. In this keynote, I will introduce the fundamental working principle of vine robots and present recent advances in mechanisms that overcome these limitations, enabling practical deployment. I will describe new approaches for high-curvature steering, modular tip-mounting of sensors and end-effectors, and efficient retraction strategies, each designed to expand the capabilities of soft growing robots. These innovations open the door to a wide range of impactful applications, from disaster response and search-and-rescue operations in collapsed structures, to directional drilling and underwater exploration, to minimally invasive medical procedures such as colonoscopy. By bridging fundamental mechanisms with practical implementation, this talk highlights how soft growing robots are transforming from laboratory prototypes into versatile tools for some of society’s most urgent and delicate challenges.

The endurance of AUV is closely related to its sailing resistance, the amount of carrying batteries and equipment load, which involves interactions among multiple disciplines. In order to develop the conceptual design of a long-range AUV in the early stage, a multidisciplinary optimization design framework is presented for decision-makers to explore the given design space, which takes into account the coupling between the disciplines of hull form, structural design and energy use. A Self-adaptive Surrogate Ensemble (SASE) method is proposed to replace the expensive process of hydrodynamic analysis, a recommended approach by the China Classification Society (CCS) specification is applied to carry out the design of metallic pressure hulls, and the classical lamination theory and Tsai–Wu criteria are adopted in the design of composite pressure hulls. The evaluation model of AUV endurance is created from the perspective of energy capacity and consumption. The conceptual design of a 200 kg-class AUV is executed to maximize the endurance based on the proposed multidisciplinary optimization design framework. Then, a brief introduction is given to the key technical issues involved in the development of long-range autonomous underwater vehicles. Finally, The Sea-Whale 2000 AUV was developed based on the optimal result and the excellent endurance performance in the sea trial validated the effectiveness of the proposed design method.

Human-robot complementary collaboration can leverage both human and robotic strengths in handling manipulation tasks in unstructured and uncertain environments effectively, either in workspace involving both human workers and robots or in hazardous environments for humans. Important challenges include interaction and perception modalities and control methods to best utilize human guidance and autonomous robotic actions. I’ll talk about the research work for side-by-side human-robot direct cooperation and human-robot collaboration in tele-manipulation at WPI and their applications. I’ll also briefly introduce human-robot social interaction.

This talk will explore a number of approaches to designing and fabricating robots that can robustly interact with the environment through embodied intelligence. This involves developing and exploiting materials, structure and sensory-motor control, to provide robots with advantageous capabilities. These approaches stem from bio-inspiration and biomimicry but also exploring computational approaches to design. The applications and new capabilities enabled by these robots will be discussed, with a focus on sustainability and agricultural applications.

Self-driving laboratories are advancing rapidly, enabling high-throughput and reproducible experiments through the integration of AI and automation. While these systems accelerate existing workflows, the challenge remains to expand scientific exploration into domains that exceed human cognitive and physical capacities. Our vision is to go beyond mimicking human scientists by developing AI-robotic systems capable of conducting experiments that humans alone cannot perform. This requires the integration of cognitive AI for hypothesis generation, physical AI for delicate and adaptive manipulations, and mathematical frameworks for their seamless coordination, supported by innovations in robotic hardware. Through this approach, we have developed prototypes of AI-robot scientists and applied them to real scientific contexts, ranging from plant biology to animal and human-related studies, generating new discoveries and insights. This presentation introduces the concept and achievements of this transdisciplinary project, highlighting how the co-evolution of humans and AI-robots can open unprecedented avenues for scientific discovery.

Transparent decision-making enables humans to understand, interpret, and predict what robots do. Interpretable and explainable methods enhance transparency: interpretable methods clarify how a learned model reaches decisions, while explainable methods articulate why specific decisions were made. In this talk, I will first introduce our interpretable AI methods that generate compact, general semantic models to infer human activities, enabling robots to gain a high-level understanding of human movement. Next, I will present our causal approach, which enables robots to rapidly predict and prevent both immediate and future failures, helping them understand why failures occur, learn from mistakes, and improve future performance. Finally, I will discuss how we combine these methods into a single framework that integrates symbolic planning with hierarchical reinforcement learning. This integration allows us to learn flexible, reusable robot policies for manipulation tasks, yielding coherent sequences of actions that can be executed independently. Interpretable and explainable AI are key to developing general-purpose robots. These approaches enable robots to make complex decisions in dynamic and unpredictable environments.

"With the spread of foundation models and affordable humanoid platforms, the possibilities of humanoid research are becoming increasingly tangible. As the next stage of development, this talk will address two perspectives: hardware innovation and the deepening of human-centric applications. On the hardware side, I will present a trajectory that extends from our previous work on tool-use intelligence—where robots adaptively employ tools for tasks—towards reconfigurable humanoids that can construct their own bodies according to the task at hand. In a separate line of research, I will introduce bio-inspired humanoids evolving towards tissue-level imitation, aiming to contribute to scientific understanding by mimicking biological organization at the structural level. On the application side, I will highlight the vision of human-centric humanoids that truly serve people, showcasing an example of a humanoid capable of assisting a person to stand up or be lifted. Through these discussions, I would like to explore with the audience how humanoid research can be further developed beyond the current state."

Deep learning-based artificial intelligence (AI) has shown to be a breakthrough technology in many fields including robotics, robot vision, industrial pattern recognition, and medical imaging. The performance of deep learning can exceed even human performance when it is trained with "big data". However, there are many areas where big data is not available. Thus, a chief limitation of deep learning is the requirement of "big data". My group has been actively studying on deep learning in medical imaging in the past 25 years, including ones of the earliest deep-learning models for image generation and lesion detection and classification in medicine. In this talk, "small-data" AI that can be trained with a small number of images is introduced. We applied our small-data AI to develop AI-aided diagnostic systems (“AI doctor”) and image generation for diagnosis (“virtual AI imaging”), including 1) AI systems for detection, segmentation, and diagnosis of major and rare cancers in medical images, and 2) virtual AI imaging systems for separation of bones from soft tissue in x-ray imaging and for denoising and quality improvement in x-ray imaging and computed tomography. Some of them have been commercialized via FDA and other regulatory approvals in the U.S., EU, and Japan, including the world-first FDA-approved deep-learning product. Our small-data deep-learning technology would be useful for the development of AI in “small-data” areas where “big data” are not available.

Conventional omnidirectional wheel mechanisms are limited in their ability to climb steps and cross gaps. The Omni-Ball, consisting of two connected hemispherical wheels, overcomes these limitations by enabling the crossing of higher obstacles and larger gaps than previously. By elongating the Omni-Ball longitudinally into a cylinder shape, we obtained the Omni-Crawler, which enables omnidirectional mobility on rough terrain. In addition, transforming the cylinder shape into a torus with inner-outer membrane motion not only enables robotic mobility in murky water, but makes it possible to further transition from Omni-Crawler to Omni-Gripper. Conventional soft grippers are not suitable for objects with sharp sections such as broken valves and glass shards, but the torus shape solves this problem by using a three-layered variable stiffness skin-bag made of cut-resistant cloth. A similar function could also be achieved using a string of beads made of titanium which can grip objects of almost any shape, even when they are on fire. To build on these gripper mechanisms from the viewpoint of bioinspired robotics, we also developed a structure inspired from the proboscis (mouthpart) of Nemertea, also known as the ribbon worm, and combined it with self-healing materials to realize a robotic blood vessel with active self-healing properties. Through the addition of repair mechanisms, we expect it to be possible to achieve the active transformation of one’s own body, thereby creating the ultimate robotic mechanism. Thus, the perspective of topology can be harnessed in the design of robotic mechanisms, culminating in the establishment of a new academic discipline—Topological Mechanism Science—as a counterpart to topological geometry.

Flapping-wing flight at the insect-scale is incredibly challenging. Insect muscles not only power flight but also absorb in-flight collisional impact, making these tiny flyers simultaneously agile and robust. In contrast, existing aerial robots have not demonstrated these properties. Rigid robots are fragile against collisions, while soft-driven systems suffer limited speed, precision, and controllability. In this talk, I will describe our effort in developing a new class of bio-inspired micro-flyers, ones that are powered by high bandwidth soft actuators and equipped with rigid appendages. We constructed the first heavier-than-air aerial robot powered by soft artificial muscles, which can demonstrate a 1000-second hovering flight. In addition, our robot can recover from in-flight collisions and perform somersaults within 0.10 seconds. I will also discuss our recent progress in incorporating onboard sensors, electronics, and batteries.

Due to their continuum structures, an existing challenge in soft robotics is creating sensors for proprioception and exteroception to facilitate control and reconfigurability. I will discuss some sensing-related challenges in these soft applications and present recent work that applies these concepts to origami robots, grippers, and wearable devices. I will also present work in enhancing the stability and mechanical selectivity of stretchable sensors and discuss applications for such sensors in wearable healthcare applications, soft robotics, and beyond.

Soft wearable robots promise to support, assist, and augment human movement in ways that rigid machines cannot. Yet building technology that works with the complexity, variability, and softness of the human body remains a significant challenge—both technically and conceptually. In this talk, I will share our experiences at the Soft Robotics Research Center at Seoul National University, where we have pursued this challenge through long-term, interdisciplinary research. Our efforts span soft wearable robots for various parts of the body, human intention detection, soft sensors, and wearable system design. Key results include multimodal sensors for activity recognition, vision-based intention detection, and wearable robotic devices for the hand, back shoulder and ankle support. These outcomes came from the collaboration between engineering, medicine, biomechanics, and design in creating systems that are not only functional, but also comfortable, and usable in everyday life. Looking ahead, the Center will become a core unit of the new SNU Robotics Institute, a university-wide initiative to unify and advance robotics research across disciplines at SNU.

Humanoid robot has potential applications in a variety of areas. However, poor locomotor energy efficiency, limited manipulation capability and poor physical human-robot interaction safety significantly hinder its advance and practical application, posing a great challenge in the robotics field. To address this problem, we propose a novel idea of bionic layagrity robotic system, inspired by the human musculoskeletal system. We reveal the fundamental principle of biological layagrity system and associated mechanical intelligences by analysing the effects of material property, morphology and topology of the musculoskeletal system on economical locomotion and versatile hand manipulation. By employing advanced functional materials and state-of-art manufacturing technologies, we finally achieve human-like locomotor system with low energy cost and bionic robotic arm-hand system with dexterous manipulation skills and excellent human-robot interaction safety. This will provide theoretical foundation and enabling design and manufacturing techniques for future advanced humanoid robotic systems.

Animals in their natural habitats exhibit extraordinary multimodal locomotion, exemplifying their remarkable capacity to effortlessly switch between different movement modes. This unique ability enables them to rapidly adapt to various environmental challenges, evade predators, and optimize strategies for capturing prey. Inspired by these biological wonders, our research seeks to advance robotic systems capable of achieving multimodal motion, designed to navigate unstructured and dynamic environments while fulfilling complex tasks. During this talk, I will showcase three compelling examples of multimodal robots that leverage soft materials and highly adaptable structures: 1) a multimodal robot engineered to cross air-water boundaries and hitchhiking on complex surfaces, 2) an octopus-inspired soft robotic arm equipped with stretchable electronics that provide bending and suction capabilities for interaction with the environment, and 3) a miniature morphable robot designed for deep-sea environment, demonstrating multiple locomotion modes. Furthermore, I will discuss several critical challenges that needs to be tackled to elevate the operational potentials of multimodal robots, ultimately paving the way for enhanced operational capabilities in unstructured and dynamically changing environments in the future.

Robotics at small scales has attracted considerable research attention both in its fundamental aspects and the potential for biomedical applications. As the characteristic dimensions of the robots or machines scaling down to the milli-/microscale or even smaller, they are ideally suited to navigating in tiny and tortuous lumens inside the human body which are hard-to-reach using regular medical tools such as endoscopy. Although the materials, structural design, and functionalization of miniature robots have been studied extensively, several key challenges have not yet been adequately investigated for in vivo applications, such as controlled locomotion of the microrobots in dynamic physiological environment, in vivo tracking, the efficiency of therapeutic intervention, biosafety of the miniature agents, and autonomy levels of the microrobotic platform. In this talk, I will first present the recent research progress in development of magnetic microrobots, from the biohybrid designs, motion control, and the rise of intelligence to rapid endoluminal delivery using clinical intervention tools. Then the key challenges and perspective of using small-scale robots, from individual to microswarms, for clinical applications with a focus on endoluminal procedures will be discussed.

A central goal in robotics is to build machines that can fluidly learn complex skills from human demonstration and interact safely and reliably in unstructured environments. While imitation learning has shown significant promise, conventional methods often struggle with a fundamental trade-off between accurately reproducing demonstrated motions and guaranteeing the stability and generalization required for real-world deployment. This talk will present a principled approach to robotic skill learning rooted in the theory of dynamical systems (DS), which models movements not as fixed trajectories, but as vector fields that guide the robot towards a goal with inherent robustness to perturbations. We will trace the evolution of DS-based imitation, from early concepts of movement primitives and Dynamic Movement Primitives (DMPs) to modern techniques that formally address the stability-accuracy dilemma. A key focus will be on the use of diffeomorphic transformations, powered by invertible neural networks, to learn complex, nonlinear skills while providing mathematical guarantees of global stability. This framework enables novel applications in high-precision tasks, such as robotic drawing and forgery detection in signatures, and extends to periodic motions for collaborative tasks like physical rehabilitation through the integration of Neural Liénard systems. Finally, the talk will explore the burgeoning synergy between classical dynamical systems and the new era of foundation models. We will discuss how DS principles can enhance modern AI, offering computationally efficient and physically grounded alternatives to Transformers, such as State Space Models, and improving the physical realism of world models through methods like Variational Information Bottleneck. Conversely, we will look to the future, investigating how large-scale models can help solve long-standing challenges in control theory, such as the automated discovery of global Lyapunov functions. This synthesis charts a path toward a new generation of robotic intelligence that combines the rigor of control theory with the expressive power of deep learning.

Recent advancements in assistive and rehabilitation robotics have demonstrated the growing potential of current technologies to restore communication with the nervous system along both afferent and efferent pathways, primarily through the development of closed-loop human–robot interfaces. This talk will explore the synergistic interaction between users and robotic systems, highlighting how such integration enables motor recovery and functional substitution in individuals with motor impairments or limb loss. Starting from a critical overview of the state of the art, the presentation will delve into recent advances in multimodal interfaces and sensory feedback mechanisms, including haptic and proprioceptive feedback, integrated into rehabilitation and assistive robots. Finally, the discussion will open toward future perspectives, highlighting key challenges and research directions in the field, including long-term adaptability, user-specific customization, and the convergence of bioengineering, AI, and robotics to shape the next generation of assistive and rehabilitative technologies.

In the dynamic field of assistive technology, soft wearable exosuits represent a significant breakthrough, setting them apart from traditional rigid exoskeletons. However, the complexity of mastering soft structures is significant: it involves not just handling the non-linear dynamics of the device but also accurately interpreting the physiological signals that are crucial to the exploit a human control loop control. My talk will cover the latest advancements from my team over the past five years, detailing our development of compact, robust, reliable, and efficient exosuits. I will discuss the critical role of integrating biomechanical modelling into control strategies to customize how the machine interacts with the user’s biomechanics, aiming to enhance human performance in tasks like collaborating with industrial manipuilators or improving running endurance. I will also introduce a new method called 'Context Aware Control,' which combines traditional control techniques with machine learning, including artificial vision, to fine-tune the assistance provided. This approach endows our exosuits with the unique ability to adapt to varying external conditions or environmental changes, significantly improving the user’s integration with these wearable robotic systems.

Advances in multi-agent systems (MAS) for applications such as surveillance and environmental monitoring demands a paradigm shift from performance optimization to holistic safety assurance. This talk presents a safety framework that addresses critical threats across both cyber and physical domains to enable reliable autonomous self-deployment. We first outline foundational cybersecurity strategies focused on privacy preservation and network resilience against data theft and service disruptions. We then turn to constrained coordination, emphasizing collision avoidance, handling communication delays, and maintaining connectivity to ensure physical and operational safety. By integrating cybersecurity with constrained coordination, the talk offers a unified approach to safety-aware MAS design. We conclude with future directions for extending these principles to heterogeneous and energy-constrained systems.

The growing demand for food, combined with labor shortages and the need for sustainable practices, is driving a profound transformation in agriculture. Under the banner of Agriculture 4.0, digital technologies, automation, and data-driven decision-making are reshaping the way we produce food. The integration of robotics into agriculture is foreseen as a key enabler of this shift, offering new ways to monitor, manage, and optimize farming systems. This talk explores recent advances in perception for agricultural robotics, with a focus on how SLAM and vision-based methods support crop monitoring and precision farming practices. I will discuss key challenges such as dealing with unstructured environments, seasonal variability, and plant occlusions, and highlight opportunities for combining multi-modal sensing.

The best current robots still fall short of the efficiency and safety guarantees exhibited by humans. One way to understand this superior performance is to develop computational models that predict how humans select, execute, and learn everyday movements like legged locomotion. Despite this need, most of our current computational and theoretical understanding of human movement is limited to simple tasks or explanatory models with limited predictive breadth. My talk will highlight the predictive principles of safe and efficient locomotion we've uncovered recently: the cost functions, controller structures, and learning rules. These principles will provide a blueprint for engineering human-like performance in wearable and autonomous robots.

"Recent years have seen remarkable progress in legged robotics, with quadrupeds and humanoids now demonstrating athletic behaviors that were out of reach only five years ago. In parallel, actively powered lower-limb prostheses have advanced rapidly, with open-source platforms such as the Open-Source Leg broadening access and accelerating innovation. Yet controlled tests in the lab only go so far. The variability of real-world environments and human users presents challenges that cannot be fully anticipated during development. To confront this gap, the first part of the talk will highlight recent work on controlling the MIT Mini Cheetah, focusing on computational methods that enable the robot to reason about its actions on the fly in novel environments. The second part will present ongoing research on improving user interfaces for lower-limb prostheses, aiming to make human–robot interaction more seamless and intuitive. Together, this work lays a foundation to expand the versatility of robotic systems in open worlds, paving the way for broader adoption in the "wild".

"Robot intelligence does not emerge from data alone. Much like humans, robots can be instructed through explicit teaching or learn by imitation. Yet the most profound form of learning arises through experience, through acting, sensing, and adapting in the world. To build robots that truly learn, we must give them the capacity to generate their own data through physical interaction.                   In this keynote, I will discuss how equipping robots with advanced sensors, compliant morphologies, and artificial skins can transform their bodies into perceptive surfaces. These designs enable robots to explore, adapt, and interact safely with humans. Unlike pre-collected datasets, this data is grounded in physical experience: robots bump, grasp, yield, and recover, constructing their own understanding of themselves and their environments.                   Such sensorized and adaptive bodies make it possible for robots to continuously gather the experiential data that supports learning while ensuring safety in human–robot interaction. In this emerging paradigm, the body is not just a container for sensors, it is the generator of data, the mediator of safe interaction, and the foundation of robotic intelligence."

Touch sensing is a crucial sensing modality in humans. However, compared to vision and audio, touch sensing is still in its infancy -- both in scientific knowledge and in applications. This talk will present an overview of recent developments in touch sensing: from new superhuman hardware, to the development of new AI algorithms for touch processing, and demonstrations of the importance of touch in robotics. Finally, I will discuss open-challenges and opportunities in this growing field.

It is increasingly recognized that less than 10% of human neural activity is consciously accessible, while the vast majority is processed unconsciously. Importantly, such unconscious processing is not merely reflexive but reflects a sophisticated form of “unconscious intelligence,” which underlies adaptive motor control and contextual meaning attribution, closely related to the concept of affordances. Recent findings indicate that unconscious intelligence does not emerge from global optimization of brain activity but rather from the integration of localized processes. While adaptive in many cases, this mechanism can also result in maladaptive states, giving rise to brain-generated disorders. Notable examples include chronic nociplastic pain and functional movement disorders, where symptoms such as persistent pain or impaired voluntary movement occur in the absence of significant structural abnormalities. These conditions, which I describe as “software diseases,” are notoriously difficult to treat within conventional paradigms focused on hardware-level pathology. Our research aims to visualize maladaptive processes of unconscious intelligence and to develop novel therapeutic strategies by bridging conscious and unconscious functions. Building on recent insights that unconscious processes can, at least partially, be accessed via associative mechanisms, we are developing Awareness AI—a framework that integrates conscious and unconscious intelligence to resolve maladaptive loops and to promote positive awareness, such as relief of pain and recovery of movement. To this end, we have established a high-density electromyography platform that enables functional estimation of limb muscle activity, including deep muscles reflecting unconscious control. This system has successfully identified causative muscles in patients with writer’s cramp and other functional movement disorders, facilitating effective recovery by targeted intervention below the threshold of conscious awareness. These advances demonstrate the feasibility of restoring adaptive function through precise measurement and intervention at the unconscious level. Through Awareness AI, we seek to establish a new medical paradigm in which software diseases are addressed by harmonizing conscious and unconscious intelligence, ultimately contributing to innovative strategies for the treatment of brain-generated disorders.

Robots for extreme environments and complex task are challenged in both the mechanical and electronic systems. Natural creatures may inspire the soft material and system design for excellent adaptability and multi-functions. This talk focuses on the interdisciplinary research between mechanics and soft robotic system. We discuss the mechanics guided strategy for extreme conditions, with the following 3 examples, soft robot in deep sea, soft hydrogel in extreme temperature and mechano-intelligent knots for surgery robot suture: 1) The design theory of deep sea soft robot intelligent system is developed, as well as its manufacturing and application methods. Put forward the principle and control method of soft robot system adapting the deep sea pressure. For the first time, the soft robotic fish operate in the Mariana Trench at the depth of 10900 m and free swim in the South China Sea at the depth of 3224 m. (Nature 2021,591, 66-71); 2) For the soft hydrogel in extreme temperature, we introduce a strategy called “hydro-locking,” which involves immobilizing the water molecules within the polymer networkof the hydrogel. Under the hydro-locking mode, the hydrogel remains highly watery and stretchable under the temperature from –115° to 143°C.  The strategy may inspire future soft robot and device design. (Science 2025, l387, 967-973); 3) Through topological design, slipknot tying and releasing can encode and deliver “mechano-intelligent” force transmission in operations and robotic suturing, which previously need force control sensors. Furthermore, the force encoding and releasing mechanism of slipknots is validated in micro structure and heavy-duty robotic operations, spanning wide geometric and force scales.

The convergence of technologies—from foundation AI models to diverse sensors and actuators to ubiquitous connectivity—is transforming the nature of interactions in the physical and digital world. People have accelerated their collaborative connections and productivity through digital and immersive technologies, no longer limited by geography or language or access. Humans have also leveraged and interacted with AI in many different forms, with the advent of hyperscale AI models (i.e., large language models) forever changing (and at an ever-astonishing pace) the nature of human-AI teams, realized in this era of the AI “copilot.” Similarly, robotics and automation technologies now afford greater opportunities to work with and/or near humans, allowing for increasingly collaborative physical robots to dramatically impact real-world activities. It is the compounding effect of enabling all three capabilities, each complementary to one another in valuable ways, and we envision the triad formed by human-robot-AI teams as revolutionizing the future of society, the economy, and of technology.

Humanoid robots are poised to become a transformative technology, with their societal and industrial roles expanding rapidly. This surge creates exceptional opportunities for both fundamental research and industry innovation in humanoid robotics. At the core of their future capabilities lies physical intelligence—the ability to sense, plan, decide, and execute motion with human-like versatility. We present two complementary foundation models as the backbone of this intelligence: a prefrontal/sensory–cortex-like model for perception, planning, and decision, and a motor-cortex–like model for precise, whole-body control. Both rely on large-scale, multimodal data, but each requires a specialized data pipeline designed to meet its unique functional and temporal requirements. The talk will highlight key industrial achievements shaping today’s humanoid landscape and present cutting-edge academic advances toward the two foundation models of humanoid physical intelligence. I will present a holistic perspective that unites model- and rule-based methods with data-driven approaches, emphasizing key research findings and supporting experimental evidence. I will conclude by highlighting the key research challenges for humanoid robots to reach their full potential in society.

Transformer-based large language models (LLMs) have achieved remarkable success across both language and vision tasks, with their impact now extending into robotics—for example, through VLA models in robotic manipulation. Despite these advances, many open questions remain. In this talk, I will focus on one fundamental question: Do all tokens require the same amount of computation within a Transformer? I will share insights into this question and present preliminary approaches to adaptive inference, in which different tokens are generated using varying numbers of Transformer layers. Actually many layers can be automatically skipped without compromising output quality. The overarching goal is to demonstrate how such methods can enhance the efficiency of Transformer-based models and improve their applicability to domains beyond LLMs.

This presentation will briefly cover recent technological developments and achievements of Unitree and our global customers. Furthermore, it will outline a vision for international collaboration to accelerate the growth of the intelligent robotics industry in the next few years.

Intelligent robotic systems are widely applied in many areas such as inspection, search and rescue, co-manipulations in Industrial 5.0, healthcare services, and logistics etc. Robots with effective intelligent adaptive control are more efficient and with more operational capability in achieving the tasks. Effective adaptive interaction of robot-robot and human-robot interaction becomes more challenging when the robots vary in terms of hardware, size, and functionalities within dynamic environments. In this talk, I will outline the challenges of the navigation and control of intelligent robotics working in unknown and dynamic environments and will present on several recent innovative intelligent adaptive control approaches we have verified through experimental studies. Specifically, results on the vision-based motion planning, intelligent navigation avoiding dynamic obstacles, adaptive robust control for multiple aerial and ground vehicles, adaptive dexterous manipulations interacting with human, and adaptive cooperative manipulation systems will be presented. The robot system is to dynamically adapt to the environment through intelligent planning and adaptive control, avoid obstacles and prevent collisions during the mission. While interacting with human, sensor-based learn-from-demonstration and adaptive admittance control grant the system a level of compliance for safe human-robot physical interaction.