Humanoid robotics in the era of foundation models

Abstract: Recent advances in foundation models have shown great promise in imitating teleoperation demonstrations for complex manipulation tasks. These models are built on a vision-language model (VLM) pre-trained on large-scale internet (non-robot) data and are connected to an action module that maps the output of the VLM to robot actions. While very successful, current methods mostly focus on static manipulation problems and fall short in providing a scalable path towards general-purpose humanoid loco-manipulation, mainly because it is impractical to generate a large amount of tele-operation demonstrations for humanoid robots. In my talk, I outline three main developments that, when combined, provide a scalable framework for generating large-scale data required for training humanoid VLAs. The first component is a general optimization-based task and motion planning (TAMP) framework that generates diverse strategies for achieving different tasks. To enable fast and efficient search, second component uses pre-trained VLMs to provide various manipulation sequence proposals (subgoals) given a desired task and environment. These two components together generate a detailed interaction graph between the robot and the environment. Finally, the third component is a generalist RL policy trained to realize any given desired interaction sequence and enable robust sim-to-real execution of the generated behaviours on the real robot.

Bio: Majid Khadiv is an assistant professor in the school of Computation, Information and Technology (CIT) at TUM. He leads the chair of AI Planning in Dynamic Environments and is also a member of the Munich Institute of Robotics and Machine Intelligence (MIRMI). Prior to joining TUM, he was a research scientist at the Empirical Inference Department at the Max Planck Institute for Intelligent systems. Before that he was a postdoctoral researcher in the Machines in Motion, a joint laboratory between New York University and Max Planck Institute. Since the start of his PhD in 2012, he has been performing research on motion planning, control and learning for legged robots ranging from quadrupeds, lower-limb exoskeleton up to humanoid robots.

Toward Human-Humanoid Cooperative Mobility: Control and Physical Intelligence

Abstract: Humanoid robots are increasingly envisioned as capable partners for physical assistance in everyday human mobility tasks. Among these, the sit-to-stand-to-sit transition remains a particularly challenging whole-body interaction problem, requiring the robot to coordinate contact forces, balance, motion generation, and human-intention understanding in real time. I will present the key algorithmic and control ingredients that enable a humanoid robot to physically assist a human, e.g., in performing sit-to-stand-to-sit motions. We detail the integration of integrating a human mode as part of the whole-body prioritized control, compliant interaction through momentum and force regulation, estimation of human states and intentions, and the modulation of support forces during critical phases of the movement. We also discuss how these components collectively create “physical intelligence” at the human-humanoid interface, enabling adaptive and safe assistance without predefined motion scripts. I will conclude with insights on generalizing these principles toward broader human mobility assistance, shared balance control, and future humanoid capabilities in physical human-robot interaction.

Bio: Abderrahmane Kheddar received his B.S. in Computer Science from the Institut National d’Informatique (now ESIE), Algiers, in 1990, and his M.Sc. and Ph.D. in Robotics from Université Pierre et Marie Curie (Paris 6), now Paris Sorbonne University in 1993 and 1997, respectively. He is Directeur de Recherche at the Centre National de la Recherche Scientifique (CNRS). He created and was the Director of the CNRS-AIST Joint Robotic Laboratory (IRL 3218) in Tsukuba, Japan (2008 - 2021). In France, he created and led the Interactive Digital Humans team at LIRMM, University of Montpellier (2010-2020); since 2020, he is leading the Bionics Platform @CARTIGEN, University Hospital Montpellier. His research interests span humanoid robotics, multi-contact motion and interaction, haptics, brain-machine interfaces, and human-robot collaboration. Professor Kheddar is a Fellow of the IEEE, the Asian Control Association, the AAIA, a member of the National Academy of Technology of France, and a Knight of the National Order of Merit of France. He has served as Associate Editor and Editor for leading robotics journals, including IEEE Transactions on Robotics, IEEE Robotics and Automation Letters, IEEE Transactions on Haptics, etc. He also contributes to several IEEE RAS committees and the IEEE Brain Initiative, and has chaired and co-chaired numerous major robotics conferences and workshops. He was William M.W. Mong distinguished lecturer at the University of Hong Kong in May 2024.

DLR Humanoids, from Earth to the Moon.

Abstract:  Robots are not only machines which are supposed to relieve humans from dangerous or routine work – they are also a scientific endeavour attempting to better understand human and animal motion and intelligence in a synthetizing way, by using the system analytic tools of engineering and computer science. The exploding commercial interest in humanoids in the last two years, with billions of dollars of investment and a large number of companies building such robots, is definitely a hype in the short run. On the other hand, the huge investment of resources and talents induced a transformation in robotics which is here to stay, and which will transform our society massively in the long run. From mechatronics and control standpoint, humanoids became a quite mature technology during the last years, and still, the development in this field continues at a high pace of innovation. The convergence of these developments with rapidly evolving artificial intelligence techniques has created the foundation for a new generation of cognitive, adaptive, multi-purpose machines - commonly referred to as Physical AI or Embodied Intelligence. or AI-Powered Robotics. In this talk, I will give an overview on humanoid robot developments at DLR, covering the entire bandwidth from design, control, over perception and cognition, up to several application examples in space and on earth.

Bio: Alin Albu-Schäffer received his M.S. in electrical engineering from the Technical University of Timisoara, Romania in 1993 and his Ph.D. in automatic control from the Technical University of Munich in 2002. Since 2012 he is the head of the Institute of Robotics and Mechatronics at the German Aerospace Center (DLR). Moreover, he is a professor at the Technical University of Munich, holding the Chair for "Sensor Based Robotic Systems and Intelligent Assistance Systems" at the School of Computation, Information and Technology. His personal research interests include robot design, modeling and control, nonlinear control, flexible joint and variable compliance robots, impedance and force control, physical human-robot interaction, bio-inspired robot design and control. He received several awards, including the IEEE King-Sun Fu Best Paper Award of the Transactions on Robotics in 2012 and 2014; several ICRA and IROS Best Paper Awards as well as the DLR Science Award. He was strongly involved in the development of the DLR light-weight robot and its commercialization through technology transfer to KUKA. He is the coordinator of euROBIN, the European network of excellence on intelligent robotics, IEEE Fellow and RAS-AdCom member.

The quest for human-level physical intelligence

Abstract: The field of robotics is advancing at an unprecedented pace, driven by promising applications in manufacturing, logistics, healthcare and personal assistance. Humanoid robots, in particular, are poised to take center-stage in our society in the near future. However, before they can live up to this promise, robots must be able to perform a wide range of everyday tasks with skill, adaptability and autonomy. While recent breakthroughs in machine learning and control have enabled impressive capabilities, scalable autonomy demands the ability to learn from humans, acquiring and refining skills just as a person would when training for a new job. In this talk, I will discuss our recent progress and open challenges in enabling robots to learn from the limited and often imperfect demonstrations humans can provide.

Bio: Prof. Stelian Coros is the director of the Computational Robotics Lab (CRL), a research group within the Institute for Intelligent Interactive Systems at ETH Zurich. His research bridges the fields of Computer Graphics, Robotics and Computational Fabrication. He draws insights from computer science, applied mathematics and control theory to establish the foundations for algorithms that address a variety of computational problems in robotics. Applications of his work range from studying the principles of dexterous manipulation and legged locomotion to computation-driven design for novel types of robots.

Robust, Social, and Superhuman: Advances in Humanoid Intelligence

Abstract: Humanoid robotics is rapidly evolving beyond isolated capabilities—locomotion, manipulation, and navigation—toward integrated, intelligent agents that reason about terrain robustness, whole-body morphology, and safe and socially compliant interaction within complex environments. This talk presents several complementary strands of recent work that collectively advance a vision of learning-driven, socially compliant, and morphologically augmented humanoid autonomy. I will first introduce a novel Learn-to-Teach reinforcement learning (RL) co-training framework, which unifies teacher and student policy learning, and Opt2Skill, which combines trajectory optimization with RL to enable dynamic, contact-rich loco-manipulation. I will then highlight two emerging directions: an emotion-aware social navigation framework that balances whole-body motion feasibility with pedestrian comfort, and augmenting humanoids with supernumerary robotic limbs, which explores morphological augmentation to expand the humanoid’s action space through coordinated multi-arm control. 

Bio: Ye Zhao is an Associate Professor at The George W. Woodruff School of Mechanical Engineering, Georgia Institute of Technology. He was a Postdoctoral Fellow at Harvard and received his Ph.D. degree from UT Austin in 2016. At Georgia Tech, he leads the Laboratory for Intelligent Decision and Autonomous Robots. His research interest focuses on planning, learning, and decision-making of highly dynamic and contact-rich robots. He received the NSF CAREER Award, ONR Young Investigator Program Award, Woodruff Faculty Fellow, and Woodruff School Faculty Research Award. He serves as an Associate Editor of T-RO, TMECH, RA-L, and L-CSS. His co-authored work has received multiple paper awards at ICRA and NeurIPS. 

Do autonomous agentic humanoids find the sense of self?

Abstract: Permeation of Agentic AI is in every detail of human-computer systems and changing the soft infrastructure of the society. Large Language Models trained using the vast document data can translate, rephrase, paraphrase the request, ask for missing information, suggest approaches and solutions. The psychological and emotional nuance can be computed once it is expressed by text or patterns of multimodality. The physical constraint of space and time is computable in the same framework. Humanoids are an ideal embodiment of Agentic AI because it is consistent with most data about the human, using which once trained they are universal at every situation where a human needs an agent. We have started discussion in our team on a problem of designing a computational pipeline for Humanoids as Embodied Autonomous Agentic AI. The current discussion is whether autonomy requests to understand human relationship. Then, when a humanoid is situated in the relationship, does it request the humanoid to have the sense of self? Though our discussion is still primitive, we would like to introduce our discussion. 

Bio: Yoshihiko Nakamura is Professor and Chair of Robotics Department, Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi, UAE. He is Professor Emeritus of the University of Tokyo and CEO of Kinescopic, Inc. He received PhD degree from Kyoto University. He held faculty positions at Kyoto University, University of California, Santa Barbara, and University of Tokyo before joining MBZUAI in 2024. Prof. Nakamura’s fields of research are humanoid robotics, biomechanics, human digital twin and their computational algorithms. He received King-Sun Fu Memorial Best Transactions Paper Award, IEEE Transaction of Robotics and Automation in 2001 and 2002. He is also a recipient of JSME Medal for Distinguished Engineers in 2019, Pioneer Award of IEEE Robotics and Automation Society in 2021, Tateisi Prize Achievement Award in 2022, and the Japan Academy Prize in 2025. 

From Video Understanding to Embodied AI

Abstract: Computer vision has recently excelled on a wide range of tasks from image segmentation to detailed image captioning and realistic video generation. This impressive progress now drives the emergence of new industries and yet, current methods still fall short in supporting simple embodied tasks. How to plant a tree, to assemble a chair or to arrange glassware without breaking it? Systems that can resolve such questions from visual inputs in real-world environments are expected to unlock immense potential of robotics and embodied decision making. Following this motivation, in this talk I will address models and learning methods that enable controlled video generation, physically-plausible animation and zero-shot solutions to manipulation tasks.

Bio: Ivan Laptev is a full professor of computer vision at Mohamed bin Zayed University of Artificial Intelligence in Abu Dhabi. Before joining MBZUAI he directed an INRIA research lab in Paris. He holds a PhD degree from the Royal Institute of Technology in  Sweden and a HdR degree from Ecole Normale Superieure in France. Ivan's main research interests include physically-grounded models for visual recognition, animation and robotics. He has published over 150 technical papers on computer vision and machine learning. He is an Associate Editor in Chief of IEEE TPAMI and previously served as an Associate Editor of IEEE TPAMI, IJCV and IVC. He also served as a Program Chair for CVPR'18, ICCV'23 and ACCV'24, and will serve as a General Chair for ICCV’29. Beyond his academic career, Ivan has co-founded a computer vision startup that grew beyond 300 employees. In 2017 Ivan was awarded a Helmholtz prize for significant impact on computer vision research.

Humanoid That Care: Toward Safe and Autonomous Assistance

Abstract: Humanoid robots are increasingly expected to operate in environments where precision, safety, and physical interaction are essential. This talk traces a humanoid’s journey from high-precision industrial manipulation to safe and autonomous physical assistance for frail individuals. We begin with a contact-rich manufacturing scenario, in which a humanoid robot operates an aircraft circuit-breaker panel using a task-space control and task-aware posture-planning framework. Building on these foundations, the talk transitions to applications in human care environments, where physical interaction requires not only accuracy but also sensitivity. I present methods for proprioceptive contact detection, compliant motion regulation, and multi-contact planning on the human body. Together, these capabilities enable humanoid robots to perceive contact through their own joints, adapt their motion safely. By following this progression from aircraft manufacturing to care facilities, the talk highlights how a unified control and perception framework can scale from high-precision industrial tasks to gentle, human-centered physical assistance - pointing toward a future in which humanoids can reliably support us across both technical and caregiving domains. Together, these results illuminate a path toward humanoids that truly care - robots that not only understand us, but can reach out and help.

Bio: Anastasia Bolotnikova received her Master’s degree in Computer Science from the University of Tartu, Estonia, in 2017. During this time, she was a recipient of the Skype and IT Academy Master’s Scholarship for excellent academic performance. Her master’s thesis, focused on humanoid robot control in aircraft manufacturing, was recognized with the Best Student Paper Award at the international conference IEEE CASE in 2017. She obtained her Doctorate degree in Robotics from the University of Montpellier, France, in 2021. Her doctoral research, conducted in collaboration with the industrial partner SoftBank Robotics Europe, focused on using humanoid robots for physical assistance to frail individuals. This work was recognized with the SoftBank Robotics Shanghai Innovation Prize at IEEE RO-MAN in 2018 and the L’Oréal–UNESCO For Women in Science Young Talents France Award in 2019. From 2021 to 2024, as part of her postdoctoral research with the RRL and BioRob laboratories, she led the Intelligent Assistive Environment project at the Center for Intelligent Systems (CIS) at EPFL, Lausanne, Switzerland. She is currently a permanent CNRS researcher in the Robotics and InteractionS (RIS) team at the Laboratory for Analysis and Architecture of Systems (LAAS), Toulouse, France.

Humanoids Behaving Like Humans: Social Navigation, Whole-Body Multi-Object Manipulation, and Perceptive Imitative Locomotion

Abstract: Humanoid robots share similar morphology with humans, allowing them to act and interact in human-like ways. In this talk, I will present three behaviors that enable humanoid robots to imitate human decision-making, manipulation, motion, and perception. First, I will introduce a social navigation framework that allows humanoid robots to make decisions based on human emotions and social interactions. Next, to explore the potential of whole-body manipulation, I will present a framework that enables humanoid robots to carry multiple objects simultaneously using their entire body. Finally, I will discuss our recent work on perceptive locomotion integrated with imitation learning, which allows humanoid robots to traverse rough terrains with more natural, human-like gaits.

Bio: Wei Zhu is an Assistant Professor in the Department of Robotics, School of Engineering, Tohoku University. He previously served as a Postdoctoral Fellow at the Georgia Institute of Technology. He earned his Ph.D. from Tohoku University and received his Master’s and Bachelor’s degrees from Nankai University. His research focuses on Embodied AI, with particular interests in social navigation for legged robots and humanoid loco-manipulation.