Interview Transcript
Q1: Could you tell us your full name, where and when you were born, and a little bit about your early days and education?
Jeannette Bohg: My name is Jeannette Bohg. I’m an associate professor for robotics in the Department of Computer Science at Stanford University. I was born in 1981 in what was then East Germany, behind the Berlin Wall. I grew up on a farm in a rural, culturally distinct region often called the “Venice of Germany” (Spreewald). I attended elementary school in my village and went on to high school to complete my Abitur. During my elementary school years, the Berlin Wall fell, which brought massive changes. A fun fact: I started learning English in fifth grade, and our teachers were only about six months ahead of us in the curriculum because previously everyone had studied Russian. As a child, I was interested in art, drawing, and languages, though not particularly biology or robotics initially. In ninth grade, my high school set up a computer room where I first learned programming using BASIC.
Q2: What was your undergraduate experience like, and where did you study computer science?
Jeannette Bohg: I graduated from high school in 2000 without a clear plan, but applied for computer science because it was seen as a field with a promising future despite the dot-com crash. I attended the Technical University of Dresden, which was an hour and a half from home. It was well known for engineering and provided a traditional, highly mathematical and theoretical computer science education. For the first time in my life, I actually struggled academically and had to learn how to study, but I adjusted after two years. Once I gained the freedom to choose my electives, I discovered I was motivated by computer vision, computer graphics, and artificial intelligence.
Q3: What did artificial intelligence look like when you were studying it at university?
Jeannette Bohg: Back then, AI coursework relied heavily on the classic Russell & Norvig textbook. Beyond introductory AI, the focus was primarily on pattern recognition (what we now call machine learning), covering Bayesian statistics, probability theory, and inference. Neural networks were considered esoteric and notoriously difficult, so I actually skipped the course on them.
Q4: How did you first get introduced to robotics during your studies?
Jeannette Bohg: There were no formal robotics courses at my university, but around 2002, I became a student research assistant working on an early, primitive video conferencing system for education—essentially a early predecessor to Zoom. Being inside a research lab exposed me to robotics work. For my diploma thesis (the equivalent of an undergrad and master’s thesis combined), I worked on vision-based Simultaneous Localization and Mapping (SLAM) with a monocular camera, reimplementing and extending groundbreaking 2005 work by Andrew Davison on Pioneer mobile robot platforms.
Q5: What did you do immediately after completing your computer science degree?
Jeannette Bohg: After graduating, I felt somewhat bored by traditional computer science because I lacked a clear connection to real-world applications or human interaction. I wanted to try something completely different, so I applied to an Art and Technology master’s program at Chalmers/IT University in Gothenburg, Sweden. The program was small—only 30 students per year—and brought together technical graduates with dancers, choreographers, musicians, writers, and painters from around the world to create physical and digital interactive art installations.
Q6: What kinds of art projects did you work on, and what insights did you gain from that experience?
Jeannette Bohg: We created interactive installations for annual exhibitions. For one team project, I worked with a Serbian student and a Mexican student to build an interactive box that visitors stepped into. Depending on how a user moved inside, the system shuffled and combined visual elements of different students to make a collaborative puzzle. It reflected the concept that in both art and science, we stand on the shoulders of giants and influence each other. Another memorable exercise involved one person describing an unseen object inside a box purely by touch while another drew it, demonstrating how challenging communication and team coordination can be across diverse lived experiences.
Q7: What prompted you to return to technical research and pursue a PhD?
Jeannette Bohg: After completing my second master’s degree in Sweden, I realized that making a living purely as a freelance artist felt too financially precarious. While looking for stable career options in Stockholm, a friend shared a link to the KTH Royal Institute of Technology website. I stumbled upon an job ad seeking three PhD students in robotics and computer vision. The topic resonated immediately with my prior research interests. I put together my application in three days, got accepted, and moved to Stockholm to begin my doctorate.
Q8: What was the focus of your PhD research at KTH?
Jeannette Bohg: I was advised by Danica Kragic. When I started, grasping and manipulation were relatively niche fields compared to SLAM, but I was fascinated by the challenge of giving robots human-like dexterity. I began by reimplementing a seminal 2006 learning-based grasping paper from Stanford. Finding its real-world failure modes inspired me to look deeper into how depth sensing and tactile feedback could enable robots to infer object geometry and explore environments beyond a single visual viewpoint.
Q9: What was the major turning point that shifted the robotics field toward learning-based manipulation?
Jeannette Bohg: Before 2006, grasping and manipulation relied almost entirely on geometric models and known shapes. While mathematically rigorous, these approaches failed in unstructured real-world environments with noisy camera data. The 2006 Stanford paper by Andrew Ng’s lab showed that learning-based methods could handle messy, partial information. Subsequent milestones—such as AlexNet on ImageNet in 2012, and Sergey Levine and Chelsea Finn’s 2016 paper on deep visuomotor policies—accelerated the field’s shift toward end-to-end reinforcement learning, imitation learning, and eventually large data-driven foundation models.
Q10: What was the main contribution of your PhD dissertation?
Jeannette Bohg: My thesis focused on multimodal scene understanding for grasping and manipulation. I combined RGB images, stereo vision, depth sensors (such as the original Microsoft Kinect), and tactile sensors on robot fingers. Integrating touch allowed robots to actively explore objects and overcome visual occlusions, yielding a far richer understanding of physical interactions.
Q11: Did human biology or cognitive science influence your research approach?
Jeannette Bohg: While traditional biology didn’t appeal to me, I read extensively in cognitive science, neuroscience, and human behavioral studies. I was fascinated by how humans adjust grip force and process skin sensor data. I wasn’t trying to build biologically exact computational models, but I was inspired by human capabilities. It’s incredible that a 20-watt human brain with high-latency, low-bandwidth neural hardware can perform physical tasks that state-of-the-art robots require massive data centers to attempt.
Q12: What did you do immediately following your PhD defense?
Jeannette Bohg: I graduated right before Christmas and started a postdoc on January 1st at the newly founded Max Planck Institute for Intelligent Systems in Tübingen, Germany. (I don’t recommend skipping a break!) Being employee number two in Stefan Schaal’s department taught me how to build a lab from scratch. While my PhD focused on perception, my goal during my postdoc was to learn closed-loop, reactive robot control and learning-based control. The Max Planck Institute provided incredible freedom for fundamental science without constant grant-writing or commercial pressures.
Q13: What kind of robots did you work with during your postdoc?
Jeannette Bohg: I integrated and wrote low-level infrastructure for a torso humanoid robot we built called “Apollo.” It had two Kuka lightweight arms, two 3-fingered Barrett hands, and a humanoid head with a friendly face, movable eyes, and a neck. We also had other humanoids, including a very specialized, expensive bipedal robot built by Sarcos to walk over rubble using lightweight titanium prosthetics instead of active ankles.
Q14: Is there a memorable story from your time working with these humanoid robots?
Jeannette Bohg: In 2012, we flew the Sarcos humanoid from Los Angeles to Frankfurt as a ticketed passenger aboard a commercial Lufthansa flight rather than shipping it in cargo. We dressed the robot in red Chuck Taylor sneakers and a Max Planck T-shirt, and programmed its face camera to display an LED smile whenever passengers approached. The 11-hour flight flew by because children and passengers constantly visited our row to interact with and take photos next to a real humanoid robot.
Q15: How did you design Apollo’s facial features and expressive capabilities?
Jeannette Bohg: The head shape was originally designed in collaboration with human-robot interaction researchers. We retrofitted an LED kit originally built for the iCub robot, replacing the red LEDs with custom blue ones to give Apollo a wider, friendlier smile whenever its vision system detected a person standing nearby.
Q16: What specific research projects did you conduct using Apollo?
Jeannette Bohg: We created large object-grasping databases to train robust point-cloud models. We also developed real-time visual object tracking systems using particle filters and extended Kalman filters designed to maintain object state tracking during sustained occlusions by the robot’s hands. By combining trajectory optimization with reactive tracking, Apollo could adjust its grasps in real time when objects or obstacles were unexpectedly moved.
Q17: What was the concept of “Interactive Perception” that you published during this period?
Jeannette Bohg: Together with several co-authors, I wrote a survey defining “Interactive Perception.” Traditional perception treats vision as a passive signal—taking a static image and processing it. Drawing from cognitive science, we argued that artificial perception should be active and interactive. Robots should deliberately touch, push, or shift their perspective to gather crucial physical information, such as determining an object’s weight, friction, or hidden boundaries.
Q18: How did you end up transitioning from Germany to a faculty position at Stanford University?
Jeannette Bohg: It was an unexpected turn of events rooted in mentorship. In late 2015, I attended a Dagstuhl seminar on robotic manipulation at a remote castle in Germany. During a discussion, Assistant Professor Sidd Srinivasa (then at CMU) asked what I planned to do after my “glorified postdoc.” When I said I’d probably just stay in my permanent group-leader role in Tübingen, he asked, “Why don’t you apply to be a professor?” That planted the seed. I put together applications for the US faculty hiring cycle on short notice, interviewed broadly, received an offer from Stanford, and made the move to California.
Q19: What differences did you notice between US and European academic interview processes?
Jeannette Bohg: In the US, faculty interviews spanned two or three intense days of back-to-back, intellectually stimulating conversations aimed at exploring potential research collaborations. In contrast, European interviews often felt like a formal exam before a long panel table, focusing heavily on evaluating whether the candidate met strict departmental requirements rather than exploring joint possibilities.
Q20: What was it like setting up your own laboratory at Stanford?
Jeannette Bohg: Establishing a lab felt remarkably like running a funded startup. Receiving a startup package gave me complete freedom to choose my PhD students, purchase hardware, design the workspace, and select research directions. Although the decision-making was intense, working directly with enthusiastic students to solve difficult problems was incredibly energizing.
Q21: What were some notable successes from your early PhD students at Stanford?
Jeannette Bohg: My first PhD student, Michelle Lee (a chemical engineer by training), worked on combining vision and touch representations for high-precision robotic assembly using reinforcement learning. That work won the Best Paper Award at ICRA 2018. Michelle went on to found a startup that automates repetitive wet-lab chemistry tasks using robotics.
Q22: How did the “TidyBot” project come about, and what was its impact on mobile manipulation?
Jeannette Bohg: I had always wanted to deploy mobile manipulators for household tasks like tidying up rooms. To overcome the lack of agile, commercial mobile bases, we acquired donated, hollow-backed omnidirectional platforms originally developed by Google X. We built “TidyBot,” a mobile system capable of learning personal user preferences to pick up objects and stow them away correctly. Because other research labs kept asking how to obtain similar bases, my visiting student Jimmy Woo created an open-source, affordable hardware design called “TidyBot++” using low-cost FIRST Robotics parts and 3D printing. This design has since been adopted widely by academic labs, startups, and even artistic interactive exhibits.
Q23: Looking back over your 20 years in robotics, how has the field changed?
Jeannette Bohg: The field has shifted away from modular, model-based pipelines (which separated perception, state tracking, and control) toward fully end-to-end deep neural networks. While end-to-end learning has enabled impressive real-world demonstrations, it comes at the cost of interpretability and theoretical guarantees. Robotics also tends to move in hype cycles—oscillating between model-based control, reinforcement learning, imitation learning, and now agentic AI architectures that flexibly combine specialized tools.
Q24: What do you see as the realistic near-term future for robotics applications?
Jeannette Bohg:
Household environments remain extremely unstructured and unpredictable, making consumer home deployment challenging in the immediate future. However, we are likely to see mobile manipulators assist in less rigid commercial settings—such as non-standard manufacturing or logistics tasks that do not require high speeds—or collaborate with humans to perform physically demanding lifting tasks to reduce workplace strain.