In this episode, Harish discusses his vision for building robots that can seamlessly integrate into human environments, exploring alternatives to the mainstream “scaling” approach in AI. The conversation covers his breakthrough work with Koopman operators for rapid skill learning, the CIMER framework that distinguishes between imitation and emulation, and his philosophy of using classical methods as “scaffolding” for modern robot learning.

Vision for robot intelligence. Harish envisions a future where robots possess intelligence that mirrors human capabilities, operating so seamlessly that their artificial nature becomes invisible. His dream centers on creating robots that can acquire skills, adapt to dynamic environments, and respond to humans in a fluent, natural manner. The ultimate goal is developing technology that follows Arthur C. Clarke’s principle: “any sufficiently advanced technology should look like magic”. This vision emphasizes three core dimensions that guide all research in the STAR Lab. First is efficiency - both from data and computational perspectives, mimicking humans’ remarkable ability to generalize skills to completely new situations with minimal examples. Second is self-reliance during operation, meaning robots shouldn’t require expert intervention once deployed, even if extensive offline expertise goes into their initial design. Third is translucency - not full transparency, but predictability that allows humans to anticipate robot behavior and engage in effective anticipatory planning. The approach differs fundamentally from pure scaling methodologies, instead focusing on creating robots that learn like humans rather than simply learning from humans. This philosophy drives the development of algorithms that can robustly acquire new skills efficiently while maintaining reliability across diverse, unstructured environments where traditional analytical methods would quickly become brittle.

New wine, old bottles. Harish’s research philosophy centers on leveraging classical approaches as “scaffolding” for modern robot learning systems. Rather than discarding decades of robotics knowledge or blindly applying it unchanged, his approach carefully identifies which aspects of traditional methods remain valuable and which create unnecessary limitations. The strategy involves borrowing core intuitions and design principles from classical robotics - such as energy conservation, stability requirements, and safety constraints - while replacing domain-specific limitations with learned components. This might manifest as neural networks constrained by physical principles, objective functions parameterized for unknown deployment environments, or learning algorithms regularized by theoretical insights from control theory. For example, the lab incorporates fundamental insights about the physical world that have been developed over centuries, from Aristotelian physics to modern control theory. These classical methods often come with optimality guarantees within their specific design specifications, but become brittle when operating outside those narrow parameters. By using them as structural foundations rather than complete solutions, the research achieves the best of both worlds: theoretical grounding with adaptive capability. This “structured robot learning” approach requires careful analysis to determine which structures are beneficial, which are harmful, and under what circumstances each applies. The goal is developing a science of structured learning that can guide the selection and combination of different structural elements based on the specific requirements and constraints of each robotic application.

Scaling isn’t everything. Harish presents a nuanced perspective on the scaling paradigm dominating current AI research, summarized in his provocative statement: “Scaling might be sufficient, but it’s not necessary”. While acknowledging the remarkable achievements of foundation models and their seemingly magical capabilities, he questions whether massive data and computational resources represent the only path to robotic intelligence. His skepticism stems from biological inspiration - humans and animals achieve remarkable learning efficiency without requiring internet-scale data. A robot designed to wash dishes doesn’t necessarily need to quote Seneca, yet current approaches to common sense reasoning seem to require reading the entire internet. This raises fundamental questions about whether alternative mechanisms exist for encoding intelligence and common sense into robotic systems. Beyond intellectual curiosity, practical and ethical considerations drive this perspective. Many institutions, industry sectors, and regions cannot afford the computational resources required for large-scale models. The concentration of these capabilities in well-resourced laboratories and for-profit companies raises concerns about equitable access to robotic technology. There’s uncertainty about how long companies will maintain open-source commitments when business incentives might favor proprietary approaches. This philosophy motivates the search for alternative approaches that can achieve comparable performance with more modest resource requirements. The goal isn’t to dismiss scaling entirely, but to develop parallel pathways that make advanced robotics accessible to a broader range of researchers, institutions, and applications, ensuring that the benefits of robotic intelligence can reach all of humanity rather than remaining concentrated among those with vast computational resources.

Making robotics accessible. Accessibility represents a core value threading through Harish’s research and outreach efforts, extending far beyond technical publications to hands-on education and practical deployment. His commitment to democratizing robotics knowledge manifests in multiple dimensions, from undergraduate circuit tutorials with 85,000 YouTube views to NSF CAREER Award research focused on resource-efficient algorithms. The technical accessibility challenge involves developing algorithms that don’t require PhD-level expertise to deploy effectively. Harish envisions putting code on GitHub that undergraduate students can successfully implement and use, removing the barriers that currently limit advanced robotics to elite research institutions. This philosophy directly opposes the trend toward increasingly complex systems that demand extensive computational resources and specialized knowledge. His writing approach exemplifies this accessibility commitment - papers from the STAR Lab deliberately avoid convoluted language, instead breaking complex concepts down to their principal components. The goal is making ideas seem “obvious in retrospect” by constructing clear conceptual graphs that connect new research to fundamental principles students already understand. Everything should trace back to undergraduate-level knowledge through logical building blocks. Educational outreach includes partnerships with HBCUs and minority-serving institutions, guest seminar series, and student research exchange programs. Working with Georgia Tech’s CEISMC, Harish introduces high school students to robotic learning research, expanding the pipeline of future researchers from diverse backgrounds. This comprehensive approach ensures that accessibility isn’t just about technical simplicity, but about creating inclusive pathways for the next generation of roboticists to contribute to the field.

Koopman operators. The discovery of Koopman operators’ effectiveness in robotic manipulation exemplifies the serendipitous nature of research breakthroughs. Harish first encountered this century-old mathematical concept through a random seminar, recognizing its potential relevance to robotic skill learning despite having no prior expertise in the area. Koopman operator theory offers a remarkable proposition: any nonlinear dynamical system can be represented exactly in a linear space by tracking how functions of the state evolve over time rather than the states themselves. While the original 1930s proof required infinite dimensions, modern computational approaches achieve excellent results with finite-dimensional approximations, similar to how neural networks approximate complex functions despite theoretical limitations. The breakthrough moment came when PhD student Yunhai Han implemented the simplest possible test - using second-order polynomials as the lifting function for learning dextrous manipulation skills. Expecting this basic approach to fail, Harish was shocked when Han returned the next day reporting that “it just works.” The method learned multi-fingered manipulation skills in seconds using only least squares optimization - no gradient descent, no complex neural networks, just high school mathematics. Harish’s initial skepticism led to weeks of code verification, unable to believe that such a simple structure could capture the complex nonlinear dynamics of contact-rich manipulation. The success challenged fundamental assumptions about the complexity required for robotic learning, demonstrating that sometimes the most elegant solutions hide in plain sight. This discovery launched a productive research direction combining classical mathematical insights with modern learning objectives, embodying the “new wine in old bottles” philosophy perfectly.

Research is random. Harish candidly acknowledges the inherently chaotic and unpredictable nature of research discovery, emphasizing that breakthrough moments often emerge from curiosity-driven exploration rather than systematic planning. The Koopman operators success story exemplifies this randomness - a chance encounter with an unfamiliar mathematical concept led to one of the lab’s most impactful research directions. Behind every published success lie numerous failed experiments and abandoned ideas that never see daylight. The research process involves generating many hypotheses, most of which prove incorrect or unproductive upon investigation. The key skill researchers develop over time is learning to distinguish promising directions from dead ends more quickly, allowing more efficient allocation of limited time and resources. This reality requires what Harish calls “irrational faith” - the passionate belief in ideas necessary to sustain the hard work required for breakthrough discoveries. Researchers must maintain emotional investment in their projects while simultaneously following rigorous scientific methodology that might ultimately disprove their hypotheses. The challenge lies in balancing personal attachment to ideas with the intellectual honesty needed to abandon approaches when evidence suggests they won’t work. The randomness extends beyond individual experiments to career trajectories and research directions. Serendipitous encounters with new concepts, unexpected experimental results, and chance collaborations often prove more influential than carefully planned research programs. Successful researchers learn to embrace this uncertainty, remaining open to surprising discoveries while maintaining enough structure and rigor to recognize and capitalize on genuine breakthroughs when they emerge from the chaos of scientific exploration.

Being a skeptic. Harish describes himself as naturally skeptical toward ideas and arguments while remaining trusting in interpersonal relationships. This intellectual skepticism serves as a crucial research tool, automatically generating devil’s advocate arguments whenever encountering new claims or proposals. The habit proves so ingrained that he must consciously moderate it during personal conversations to avoid inadvertently challenging everything his partner says. This skeptical mindset proved essential during the Koopman operators breakthrough, when Harish spent weeks verifying Yunhai Han’s seemingly impossible results. The natural inclination to question extraordinary claims - that second-order polynomials could learn complex manipulation skills in seconds - led to thorough validation that ultimately confirmed the discovery’s legitimacy. Without this skepticism, the team might have missed subtle bugs or experimental errors that could have invalidated their findings. The balance between skepticism and openness requires careful calibration. Excessive skepticism can prevent researchers from pursuing promising but unconventional directions, while insufficient skepticism can lead to accepting flawed results or building upon incorrect foundations. Harish advocates for maintaining intellectual humility - being willing to change deeply held beliefs when presented with compelling evidence. This skeptical approach extends to self-evaluation and the broader research community. Researchers must question their own work as rigorously as they question others’, avoiding the ego attachment that can blind scientists to their ideas’ limitations. The goal is fostering a culture where challenging assumptions and demanding rigorous evidence strengthens rather than threatens the scientific enterprise, ultimately leading to more robust and reliable knowledge.

Imitation vs emulation. The CIMER framework introduces a crucial distinction between imitation and emulation, illustrated through the analogy of learning to kick like Cristiano Ronaldo. Traditional imitation learning, exemplified by behavioral cloning, focuses on mimicking the expert’s exact movements - like trying to replicate Ronaldo’s leg trajectories precisely. However, this approach ignores fundamental differences in morphology, capabilities, and context that make exact replication impossible or suboptimal. Emulation takes a fundamentally different approach, focusing on achieving similar outcomes rather than identical behaviors. When learning penalty kicks from watching Ronaldo, a human observer doesn’t try to move their legs exactly like the expert. Instead, they observe what happens to the ball and attempt to produce similar results through their own unique physical capabilities and movement patterns. The CIMER algorithm operationalizes this distinction through a two-phase process. The initial imitation phase uses behavioral cloning to provide a rough approximation of the desired skill, similar to watching Ronaldo on television and developing a basic understanding of penalty kick technique. However, this phase alone proves insufficient because it lacks the refinement that comes from actual practice and feedback. The crucial second phase implements emulation through reinforcement learning with a task-agnostic reward function. Rather than designing bespoke rewards for each task, the system uses a universal emulation reward that measures the difference between the robot’s impact on objects versus the expert’s impact. This approach enables the robot to refine its movements through practice, learning to achieve expert-level outcomes through its own optimal movement patterns rather than slavish imitation of expert behaviors.

Future challenges. Harish identifies the scalability of structured approaches across diverse domains as a primary future challenge. While individual structures like Koopman operators or emulation rewards work well within specific contexts, the fundamental question remains whether universal structures exist that can generalize across all robotic tasks, or whether the field must accept a “bag of tricks” approach with different structures for different domains. The challenge extends beyond simply finding more structures to understanding how they can be composed and combined effectively. Many structures prove compatible - for example, CIMER incorporates Koopman operator theory alongside its emulation framework. However, each structure has breaking points and limitations that aren’t yet well understood. The field lacks systematic methods for selecting appropriate structures for novel situations or determining when to fall back on purely learned approaches. A particularly intriguing frontier involves enabling robots to discover their own structures autonomously. This meta-learning challenge would require robots to determine their own objective functions, policy architectures, and learning constraints based on task requirements and environmental conditions. While this capability seems far-fetched, it represents a natural extension of current representation learning research into the broader domain of structural discovery. The “unknown unknown” problem poses perhaps the greatest challenge. When robots encounter completely novel situations that don’t fit existing structural frameworks, how should they respond? Current approaches provide no principled method for recognizing when familiar structures will fail or for transitioning to alternative approaches. Solving this challenge may require fundamental advances in robot self-awareness and adaptive reasoning that go beyond current capabilities in machine learning and artificial intelligence.

Mental wellness strategies. Harish emphasizes the importance of maintaining perspective and emotional balance throughout the inherently stressful research process. His approach centers on dissociating self-worth from temporary feelings of intellectual inadequacy, recognizing that struggling with new concepts reflects the learning process rather than personal failure. This mindset proves crucial for researchers who must constantly venture into unfamiliar territory. Practical strategies include maintaining rich social connections and engaging in activities outside academia. During graduate school, Harish cultivated a large friend group that provided both emotional support and forced participation in non-academic activities even during busy periods. These relationships offered perspective during difficult phases and prevented the isolation that can exacerbate research-related stress. Physical activities, particularly hiking and spending time in nature, provide essential mental restoration. While parenting responsibilities have reduced his hiking frequency, Harish continues to prioritize outdoor activities when possible. The combination of physical exercise, natural beauty, and temporary disconnection from academic pressures offers a powerful reset mechanism for mental clarity and emotional balance. The support of understanding partners proves invaluable, particularly those with relevant professional expertise. Being married to a psychologist provides Harish with both personal support and professional insights into mental health maintenance. However, he emphasizes that similar benefits can be achieved through other support systems, including friends, mentors, counselors, and structured wellness programs that many institutions now provide for graduate students and faculty.

Meditation and perspective. Harish discovered meditation during the pandemic and found it transformative for managing the constant pressure of academic achievement. The practice provides a crucial counterbalance to the relentless pursuit mentality that characterizes much of research life, offering brief periods where no accomplishment or progress is required. Even five to ten minutes of simply sitting and being present can provide profound perspective shifts. The meditation practice reveals the often arbitrary nature of the goals and stresses that dominate daily academic life. From a meditative perspective, many urgent concerns appear almost comical in their constructed importance. This insight doesn’t diminish the value of research work, but rather provides healthy distance from the emotional intensity that can make temporary setbacks feel catastrophic. Harish acknowledges the challenge of integrating meditative insights into active research life. The perspective that reveals the constructed nature of academic pursuits can be difficult to maintain while simultaneously investing the emotional energy necessary for breakthrough discoveries. He describes enjoying “that wisdom for like five minutes” before returning to the passionate engagement that drives productive research. The practice also helps with physical restlessness and the constant mental activity that characterizes academic work. For someone who naturally fidgets and paces during lectures, meditation provides training in observing these impulses without immediately acting on them. This awareness extends beyond formal meditation sessions, creating opportunities throughout the day to pause, breathe, and maintain perspective amid the chaos of research life.

Advice for researchers. Harish’s primary advice for emerging researchers centers on maintaining intellectual openness and avoiding premature commitment to particular paradigms or worldviews. While initial fascination with specific approaches provides necessary motivation for entering the field, he cautions against forming rigid opinions before gaining sufficient experience across diverse research perspectives. The key insight is that a PhD represents a license to conduct research rather than a commitment to specific problem domains. Students shouldn’t feel constrained by their dissertation topics or advisor’s research focus, as the degree simply certifies their capability to think scientifically and conduct rigorous investigation. This perspective liberates students to explore broadly and develop their own research interests over time. Critical to this development is actively seeking out conflicting viewpoints and genuinely engaging with perspectives that challenge one’s current beliefs. Harish recommends attending seminars and talks by researchers working on fundamentally different approaches, asking honest questions about whether disagreements stem from emotional attachment to existing work or legitimate intellectual differences. The ultimate goal is developing the intellectual humility to change one’s mind when presented with compelling evidence. This requires emotional detachment from specific ideas and papers, viewing them as hypotheses to be tested rather than personal achievements to be defended. Researchers must remain genuinely open to discovering that long-held beliefs are incorrect and be prepared to pivot when evidence demands it. This flexibility, combined with exposure to diverse perspectives, enables the development of nuanced, well-informed research philosophies that can guide productive careers in an rapidly evolving field.

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

Harish Ravichandar is an Assistant Professor in the School of Interactive Computing and a core faculty member of the Institute for Robotics and Intelligent Machines (IRIM) at Georgia Tech. He directs the STAR Lab (Structured Techniques for Algorithmic Robotics), where his team develops structured computational frameworks that combine classical approaches with modern machine learning to improve robots’ efficiency, reliability, and generalizability.