In this episode, Sehoon explores the challenges of replicating human locomotion, the importance of emotional expression in robots, and the innovative approaches to cross-morphology locomotion. He opens up on his recent collaboration with Neuromeka and their multi-million dollar project building a new medical humanoid robot.

Grace in human locomotion. The challenge of replicating human locomotion grace in robots remains one of the most compelling frontiers in robotics. While many researchers believe quadrupedal locomotion has been largely solved, humanoid bipedal locomotion is far from achieving human-level grace and robustness. Sehoon observes that even with significant hardware advances, current robots lack inherent safety and reliability. The fundamental question revolves around whether the limitation is hardware-based—such as actuation speed and accuracy—or software-based, involving motion intelligence and control algorithms. Some researchers argue that with sufficiently fast and accurate actuation, locomotion could be solved simply by stepping toward where the robot is falling. Others point to sensing limitations, particularly proprioception, which provides limited but crucial environmental interaction data. Current approaches heavily rely on reinforcement learning with massive amounts of data, which differs significantly from how humans learn to walk. This sampling-based curse of reinforcement learning means that policies often struggle with unseen environments, leading to performance gaps when robots encounter new terrain. The challenge lies in bridging this fundamental difference between human learning efficiency and robot training requirements, seeking approaches that can achieve robust locomotion without requiring internet-scale data collection and processing.

Human learning vs.robot training. The stark contrast between human learning efficiency and robot training requirements presents one of the most frustrating paradoxes in robotics research. Sehoon illustrates this with a personal anecdote about his two-year-old daughter, who could walk up stairs relatively large compared to her body size while holding a cup in her mouth — a remarkable feat of coordination and adaptability. In contrast, reinforcement learning-trained robot policies remain far less robust despite requiring massive amounts of training data. This disparity stems from “the curse of the sampling-based approach” inherent in deep reinforcement learning. Current methods estimate policy gradients from enormous datasets, but gaps always emerge when robots encounter new terrain or unseen environments, resulting in poor performance. The fundamental limitation lies in the robot’s inability to generalize effectively from training data to novel situations, a capability that humans seem to possess naturally. The challenge extends beyond mere data efficiency to include the quality and nature of learning itself. Humans appear to learn through a combination of innate biological mechanisms, physical experimentation, and contextual understanding that current AI systems struggle to replicate. While robots can be trained to perform specific tasks well, they lack the intuitive understanding of physics, balance, and environmental interaction that allows humans to adapt seamlessly to new locomotion challenges. This fundamental gap suggests that achieving human-like learning efficiency in robots may require entirely new approaches beyond current reinforcement learning paradigms.

Transferring personality to robots. Humans possess an remarkable ability to express individual characteristics and personality through movement, while robots remain primarily task-driven, barely managing to execute basic functions like walking without falling. Sehoon believes the key to addressing this limitation lies in learning from human data to enable robots to express better and more nuanced personality traits. His research focuses on developing policies that can capture and transfer human characteristics to robotic systems, creating machines capable of more than mere task completion. The challenge becomes particularly complex when dealing with cross-morphology imitation—transferring personality and movement styles between different body types. While human-to-humanoid transfer is relatively straightforward due to similar morphologies, transferring human characteristics to quadrupedal robots or legged manipulators presents significant technical challenges. This requires sophisticated understanding of how personality manifests in movement and how those essential qualities can be preserved across dramatically different physical forms. Drawing from his computer graphics background, Sehoon explains that early approaches to personality transfer relied on manual, feature-based methods that measured energy or other quantifiable factors. Modern approaches increasingly use data-driven methods, collecting samples of actors displaying different personality traits—such as confident walking styles—and extracting style elements that can be transferred to characters or robots. These techniques often combine generative adversarial networks (GANs) and other learning-based approaches to model what’s common across different personality expressions while preserving individual characteristics.

Motion correspondence and skill transfer. Establishing reliable correspondence between human and robot movements presents complex technical challenges, particularly in ensuring safety and preventing unintended behaviors. Sehoon addresses this through careful definition of what constitutes successful correspondence and what failure modes might emerge. In simulation environments, the process involves iterative training where problematic kinks are identified and refined through policy fine-tuning. However, teleoperation scenarios prove more difficult, requiring deep understanding of human intention behind observed motions. The fundamental challenge lies in interpreting human intention and projecting it onto robot capabilities while accounting for noise in human demonstrations. This requires robots to understand not just what humans are doing, but why they’re doing it, then adapt those intentions to their own physical constraints and capabilities. One approach involves building comprehensive motion skill repertoires that define what robots can safely accomplish, then projecting all human inputs onto this safe operational space. Sehoon’s ACE (Adversarial Correspondence Embedding) paper demonstrates this concept through examples like sword fighting, where robots learn compact controllers that take small-dimensional latent space vectors as input to generate diverse motions. The system formulates a manifold of robot-achievable actions and establishes correspondence between human motions and this latent space. When humans act, the system already knows the mapping between human motion and robot capabilities, allowing robots to follow the manifold of achievable actions while imitating human intentions. This approach ensures safety while preserving the essential characteristics of the demonstrated skills.

Sehoon’s background. His research journey exemplifies the power of interdisciplinary thinking in advancing robotics. His career began with a passion for computer games, which led him to study computer animation during the early years of his PhD. The transition to robotics occurred during his senior years when he realized that physics-based control techniques learned from computer graphics could transfer effectively to real robotic systems. This expansion of research interests proved transformative, as seeing learned behaviors implemented on actual hardware provided an amazing experience that solidified his commitment to robotics research. His background spans multiple critical domains that few researchers can claim expertise in: character animation, robotics, and reinforcement learning. He acknowledges that other prominent researchers, such as Professor Sergey Levine at UC Berkeley, share similar interdisciplinary backgrounds. The computer graphics foundation provides unique insights into movement aesthetics, style transfer, and character expressiveness that purely engineering-focused researchers might overlook. This diverse background enables him to approach robotics problems from multiple angles, bringing artistic sensibilities to technical challenges and vice versa. His experience with animation and character design informs his understanding of how movement can convey personality and emotion, while his robotics expertise ensures that these concepts can be practically implemented on real hardware. The reinforcement learning component provides the technical framework for training systems that can learn and adapt these complex behaviors.

Navigating academia and industry. Sehoon’s perspective on the academia-industry divide comes from direct experience at both Disney Research and Google before joining Georgia Tech. He describes his industry experience positively, noting that Disney Research provided the perfect environment for animation and robotics research, while Google exposed him to numerous brilliant minds. However, he finds academia particularly rewarding, describing every day as full of surprises, especially when interacting with students. The academic environment provides approximately 90% academic freedom, which substantially exceeds what’s typically available in industry settings. This freedom allows researchers to pursue long-term, curiosity-driven projects that might not align with immediate commercial interests. The choice between academia and industry represents one of the most frequent questions Ha receives from aspiring researchers. His advice emphasizes that both paths offer valuable learning opportunities, but they serve different purposes and appeal to different personality types. Industry provides exposure to cutting-edge resources, practical constraints, and immediate application of research, while academia offers intellectual freedom, the opportunity to mentor the next generation, and the ability to pursue fundamental questions without immediate commercial pressure. The decision ultimately depends on individual priorities, career goals, and personal satisfaction derived from different types of work environments and challenges.

Neuromeka’s $7M research grant. His collaboration with Neuromeka on a multi-million dollar medical humanoid robot project represents a significant application of his research expertise to real-world healthcare challenges. Georgia Tech’s role focuses on high-level decision-making stacks, with Ha collaborating with Professor Jennifer Kim from the School of Interactive Computing. Their responsibility spans the complete spectrum from human-robot interaction—including understanding human requests and providing emotional support—to high-level planning for navigation and decision-making. The project’s technical architecture includes a mobile manipulation layer designed to move objects and provide assistance to nurses and patients. Georgia Tech handles the AI-driven research components while other institutions contribute control systems, hardware development, and real-world user studies. The goal isn’t to replace nurses or medical workers but to handle routine tasks like fetching medicine, allowing healthcare professionals to focus on more critical patient care activities. An interesting cultural insight Ha shares involves Korean hospital dynamics, where patients often hesitate to ask nurses for assistance because they feel sorry about bothering busy medical staff. The medical assistance robot could help address this by providing more autonomous support that patients feel comfortable requesting. This demonstrates how robotics solutions must consider not just technical capabilities but also cultural and psychological factors that influence human-robot interaction. The project represents a comprehensive approach to medical robotics that integrates advanced AI, social awareness, and practical healthcare needs into a single platform designed for real-world deployment.

Social navigation in crowded environments. Hospital environments present unique navigation challenges due to crowded conditions, time-sensitive tasks, and the need for robots to operate non-intrusively around rushing people. Sehoon emphasizes the need for near perfect social navigation when robots must operate alongside social entities in these dynamic environments. The approach combines large language models (LLMs) for high-level decision making with imitation learning and massive-scale reinforcement learning to create more intelligent systems than current commercial solutions. Current delivery robots, such as those commonly seen on university campuses, often exhibit overly conservative behavior that prioritizes safety at the expense of efficiency. His team’s goal is developing robots that understand social situations better and can assess risk more intelligently, taking appropriate detours when necessary while maintaining efficient operation. The key innovation lies in combining multiple AI approaches to create systems capable of nuanced social understanding. Rather than simply stopping when encountering obstacles, these robots could communicate intent through movement, sending implicit messages to nearby people about their navigation plans. For example, a robot moving sideways could non-verbally communicate that it recognizes human presence and is actively avoiding collision, allowing people to predict robot behavior and coordinate their own movements accordingly. This type of sophisticated social navigation could significantly improve robot integration in complex human environments like hospitals.

Building trust with robotic guides. The development of robotic guide dogs for visually impaired individuals presents unique challenges in establishing and maintaining trust between human users and artificial assistants. Sehoon’s research team conducted comprehensive user studies published at the Human-Robot Interaction (HRI) conference, examining communication dynamics between robotic guide dogs and the blind and visually impaired (BVI) population. Their findings revealed that different users react very differently to various forms of communication, whether verbal, haptic, or through physical pulling, with communication flowing bidirectionally between humans and robots. Unlike biological guide dogs, robotic systems could potentially explain their decision-making processes, telling users “I made a detour because of this specific reason”. This verbal communication capability could help build trust by providing transparency in robot decision-making that biological counterparts cannot offer. The trust-building process likely evolves over time as users become familiar with their robotic guides’ capabilities and reliability. Initial apprehension would naturally occur when someone depends entirely on a robot for navigation safety, but Ha remains optimistic that robotic guide systems can develop meaningful trust relationships with users. The key lies in consistent, reliable performance combined with clear communication about robot capabilities, limitations, and current situational assessments. This transparency could ultimately provide advantages over biological guide dogs, as users gain better understanding of their guide’s decision-making processes and operational parameters.

Selecting promising students. Sehoon acknowledges that student selection represents a challenging aspect of academic management, expressing curiosity about how other professors approach this critical decision. His primary indicators focus on research experience in relevant topics, particularly publication records in venues where he frequently publishes, such as ICRA, RAL, and SIGGRAPH conferences. Technical skill sets prove equally important, with preference for candidates who understand motion imitation learning, scalable reinforcement learning, and traditional legged locomotion theory. Beyond technical qualifications, Ha values direct interaction through classroom teaching, where he can observe personality traits and knowledge depth firsthand. For master’s and undergraduate students interested in joining his lab, he typically recommends taking his class first to facilitate mutual evaluation. This doesn’t require excelling in the course or waiting until semester completion—the goal is simply getting to know potential lab members better, making Ha more comfortable with his selection decisions. Personality assessment focuses on active participation and question quality during classes, which reveals how well individual students understand lecture material. When other qualifications are comparable, Ha prefers bright personalities — energetic, positive individuals who engage actively with the material and lab community. However, he emphasizes that personality preferences aren’t absolutely necessary for success. The approach balances technical competence with interpersonal compatibility, recognizing that successful research requires both individual capability and effective collaboration within the lab environment. This comprehensive evaluation process helps ensure that selected students can both contribute meaningfully to ongoing research and benefit from the lab’s educational and research opportunities.

Humanoids running fast. The question of whether humanoids can achieve elite human running speeds requires careful definition of what constitutes a “humanoid” robot. Sehoon explains that legged wheels—wheels with leg-like appendages—could easily surpass human running speeds, but whether such systems qualify as humanoid remains debatable. The continuum extends from purely biological human forms to increasingly mechanical solutions, with specialized designs incorporating springy prosthetics potentially achieving superhuman speeds. From a hardware perspective, achieving high-speed humanoid running requires energy-efficient mechanisms capable of storing and releasing energy from foot contact within very short timeframes. This demands either sophisticated energy recovery systems or significantly more powerful actuation than currently available. Software solutions must maximize energy efficiency while minimizing unnecessary dissipation—avoiding ground stomping in favor of smooth center-of-mass maintenance within minimal height windows. Ha believes the primary bottleneck for high-speed running in perfect conditions lies in hardware rather than software. For single-task scenarios like running on ideal tracks, developing highly optimized controllers through reinforcement learning, imitation learning, or both approaches seems achievable. However, hardware capabilities currently limit what’s possible despite advances in control algorithms. The reliability question compounds this challenge—while robots might achieve high speeds on perfect tracks, maintaining performance on uneven terrain or during unexpected circumstances remains problematic. Safety considerations become critical when robots operate at high speeds, as failures could result in expensive damage, engineering time loss, and potential harm to surrounding people and property.

Developing fault-tolerant robots. Creating fault-tolerant robots requires implementing sophisticated self-awareness mechanisms that enable machines to understand their operational limits and respond appropriately to potentially dangerous situations. Sehoon emphasizes the need for “safety-aware reinforcement learning” where robots must know what they can accomplish and take alternative actions when problems arise rather than blindly pursuing current tasks. This involves recognizing physical stress indicators, such as excessive knee strain, and switching to safer operational modes when such conditions are detected. The concept extends beyond simple failure detection to include capability assessment during operation. For example, if a robot attempts to lift an object that exceeds its capacity, it should recognize this limitation and either request assistance or abandon the task rather than risking damage to itself or surroundings. This requires sophisticated sensing and reasoning capabilities that can evaluate real-time physical feedback against known operational parameters. Safety awareness must be built into the fundamental control architecture rather than added as an afterthought. Robots need continuous self-monitoring systems that can detect anomalous conditions and gracefully degrade performance or switch to safe operational modes when necessary. This might involve maintaining multiple operational strategies for different risk levels, allowing robots to adapt their behavior based on current conditions and confidence levels. The ultimate goal is creating systems robust enough to operate safely in unpredictable real-world environments while maintaining useful functionality even when operating at reduced capacity due to safety constraints.

Expressiveness in robotics. Sehoon’s primary research interest centers on developing expressive controllers that enable robots to communicate emotions and interact meaningfully with their surroundings, moving beyond simple utilitarian task completion to more nuanced behavioral repertoires. This work connects directly to his early research questions about empowering robots with characteristic behaviors and personality traits. Humans excel not only at useful activities like carrying boxes from point A to point B but also at expressing emotions and engaging with their environment in sophisticated ways. The challenge lies in translating human expressiveness to robotic systems that possess different morphologies and capabilities while maintaining the essential emotional and communicative content. This requires understanding how humans use movement, gesture, and positioning to convey internal states and intentions, then developing methods to map these concepts onto robotic embodiments. The research draws from Ha’s computer graphics background, where character animation has long grappled with similar challenges of conveying personality and emotion through movement. Developing expressive humanoid controllers represents a natural extension of Ha’s work on cross-morphology imitation, social navigation, and human-robot interaction. The Neuromeka medical robot project provides an ideal testing ground for integrating these concepts, as hospital environments require robots that can communicate empathy, reassurance, and professional competence through their behavior. Ha’s approach involves combining all the techniques his lab has developed over the years—cross-morphology imitation, social navigation algorithms, and human-robot interaction principles—into comprehensive systems that can express appropriate personality traits for their intended applications. This represents a holistic approach to robotics that prioritizes emotional intelligence alongside technical capability.

Favorite researcher. Sehoon identifies his PhD advisor, Dr. Karen Liu (currently at Stanford), as his most significant influence, though he acknowledges having too many researchers he admires. From Dr. Liu, he learned not just research methodology from “A to Z” but, more importantly, the fundamental lesson that research is a fun thing. This perspective shapes Ha’s entire academic philosophy and explains his continued commitment to academia despite opportunities in industry. Dr. Liu’s consistent excitement about research meetings and new research ideas, combined with her passionate approach to exploration, provided a model that Ha naturally absorbed during his graduate training. This enthusiasm became central to Ha’s own research identity—he pursues research because he genuinely enjoys studying new topics, exploring uncharted territories, and communicating discoveries with students. The lesson transcends technical knowledge to encompass the emotional and motivational foundations necessary for successful long-term research careers. Ha describes this influence as the most valuable lesson he learned during his academic development. The emphasis on research as inherently enjoyable rather than merely professionally necessary creates a sustainable foundation for dealing with the inevitable challenges and setbacks that characterize academic careers. This perspective enables Ha to maintain genuine enthusiasm for discovery and learning, which translates into more engaging interactions with students and more creative approaches to challenging research problems. The influence of a passionate advisor demonstrates how academic mentorship extends far beyond technical training to include fundamental attitudes toward intellectual curiosity and professional fulfillment.

Advice for aspiring researchers. Sehoon’s primary advice to emerging researchers focuses on staying current with rapidly evolving research landscapes, acknowledging this as an area where he feels he could have performed better during his own development. When he was a PhD student, catching up with state-of-the-art computer animation research required reading approximately 30 papers annually covering all top conferences and journals in the field. Today’s research environment presents vastly more complexity, with tons of things going on and many interesting works. Ha observes that his current students naturally stay updated through diverse information channels including X, YouTube, personal conversations, and comprehensive paper searches. He emphasizes that understanding current developments in the research community is very important, even when specific advances may not immediately relate to one’s current research focus. This broad awareness helps researchers grow and develop more comprehensive understanding of their fields. The advice reflects a fundamental shift in how researchers must approach professional development in an era of accelerated publication rates and expanding research domains. While specialization remains important, Ha suggests that broad awareness of community developments provides essential context for individual research contributions and helps identify unexpected connections between seemingly disparate areas. This approach requires active engagement with multiple information sources and continuous learning habits that extend throughout research careers. The goal isn’t necessarily mastering every development but maintaining sufficient awareness to recognize relevant advances and potential collaboration opportunities as they emerge.

On the go? There’s an audio-only version too. Click here.

Sehoon Ha is one of the foremost authorities on learning based legged locomotion in robots. Before joining Georgia Tech as an assistant professor, he was a research scientist at Google and Disney Research. He’s interested in character animation, robotics, and artificial intelligence.