Lex Fridman PodcastMarc Raibert: Boston Dynamics and the Future of Robotics | Lex Fridman Podcast #412
CHAPTERS
Marc Raibert’s origin story: from basement tinkering to the MIT AI Lab (1974)
Marc describes growing up as a hands-on builder and the moment he became captivated by robotics: seeing a disassembled robot arm in an MIT lab. He contrasts his initial path in neurophysiology with the appeal of robotics and AI as a more “conceptual” way to study intelligence.
- •Early maker roots influenced by his father’s engineering interests and home workshop
- •1974 turning point: seeing a robot arm in pieces at MIT sparked a robotics career
- •Shift away from neurophysiology toward robotics/AI as a path to understanding control and intelligence
- •Early reflections on the brain-science vs robotics approaches to intelligence
Why dynamic, ‘aggressive’ locomotion beats cautious walking—and what it implies for manipulation
Raibert explains his dissatisfaction with early slow, statically-stable walking robots and why animals rely on balance, prediction, and energy recycling. He extends this philosophy to robot manipulation, arguing robots must become more dynamic and less ‘safe’ to reach human-like dexterity.
- •Static tripod-stability walking is unlike animal locomotion; real movement requires balancing and prediction
- •Importance of springs/compliance and energy circulation in legs
- •Early focus on bouncing energy later expanded to balancing in 3D
- •Analogy to manipulation: humans nudge/roll/juggle rather than only careful ‘grasp-and-move’
The first hopping robots: funding hustle, early prototypes, and core control ideas
Marc recounts how a small seed fund from Ivan Sutherland led to a pivotal DARPA pitch—and a major grant that enabled the first serious hopping robots. He outlines the technical building blocks of pogo-stick locomotion: energy regulation, foot placement, and body attitude control.
- •Ivan Sutherland mentorship and an initial $3,000 prototype effort
- •Old-school DARPA hallway pitch leads to $250k funding in 1980-era dollars
- •Hopping control decomposed into: bounce energy, foot placement relative to COM, and body attitude stabilization
- •Progression from constrained planar experiments to full 3D balancing
From CMU to the (MIT) Leg Lab: scaling from one leg to quadrupeds
Raibert situates the Leg Lab’s beginnings at Carnegie Mellon and tracks the timeline from early 1980s hopping machines to 3D locomotion and quadrupeds. He emphasizes that breakthroughs required years of iteration and the right collaborators.
- •Leg Lab started at CMU (1980–1986), later associated strongly with MIT
- •Milestones: simplified hopper (~1982), 3D hopper (~1983), early quadruped (~1984–86)
- •Iterative development: years to turn concepts into robust machines
- •Finding great people (e.g., Ben Brown) was a key leverage point
Founding Boston Dynamics: simulation-first, then a robotics ‘return’ via Sony
Marc explains Boston Dynamics’ early identity as a physics-based simulation company, not initially a robot-product company. Work with Sony (AIBO Runner, Qrio tools) pulled the team back into embodied robotics and reset the company’s direction.
- •Boston Dynamics founded in 1992 with simulation as the initial core
- •Sony collaboration: modified AIBO legs to create ‘AIBO Runner’; also work on Qrio tooling
- •Pre-Zoom era remote collaboration (ISDN telecons) with Japan-based teams
- •Realization: robotics is where the company ‘belonged’
Early BD experiments and the pivot lesson: the surgical simulator that didn’t fit
Raibert shares a vivid story of building an advanced force-feedback surgical training simulator and showcasing it successfully—only to discover there was no viable business path for a tiny bootstrapped company. Killing the project clarified BD’s identity and sharpened focus.
- •Built a mirror-based surgical simulator with force feedback, 3D graphics, and scoring
- •Trade show success revealed surgeons’ competitiveness—but not a sustainable revenue model
- •Bootstrapped constraints made hospital sales/marketing unrealistic
- •Decisive pivot: abandon even fascinating tech that doesn’t match company capabilities
BigDog: DARPA’s biodynotics, field testing in mud, and full onboard integration
BigDog becomes the defining milestone that scaled Boston Dynamics and validated rugged legged mobility. Marc contrasts earlier lab-bound robots with BigDog’s fully integrated power/computation and the ‘build it, break it, fix it’ culture forged in real-world trials.
- •DARPA biodynotics program catalyzed BigDog; BD one of few funded proposals
- •Key hires (e.g., Martin Buehler) helped move from lab demos to outdoor field testing
- •Earlier quadrupeds relied on offboard hydraulics and computers; BigDog integrated engine power and onboard systems
- •Quantico trail testing (Guadalcanal Trail) as a demanding real-world milestone
From BigDog to LS3 and Spot: autonomy lessons, operator roles, and scaling down
Marc discusses the realities of early operations—human operators provided perception while the robot handled balance and foot placement. He traces the lineage BigDog → LS3 and the later push toward a smaller, less intimidating robot that helped define Spot’s form factor and usability.
- •Real-world operation initially required skilled human teleoperation for perception and navigation choices
- •Later electric precursors became easier to operate, even by amateurs
- •LS3 load-carrying evolution and extreme payload anecdotes (beyond design spec)
- •Larry Page’s prompt: build a ~60 lb robot suitable around people, helping inspire Spot
Hydraulics vs electric actuation: performance, stigma, and BD’s valve innovations
Raibert defends hydraulics as a high power-density technology and details Boston Dynamics’ major engineering advances in valves and compact power supplies. He frames the hydraulic-to-electric shift as partly about product fit and perception, not just raw capability.
- •Hydraulics deliver strong performance in lightweight packages, but can be messy (drips) and seen as ‘old-fashioned’
- •BD innovated hydraulic valves beyond classic 1950s aircraft designs (smaller, lighter, more efficient)
- •Compact hydraulic power unit integration (motor, pump, filters, heat management)
- •Electric platforms emerged for smaller, friendlier robots, while hydraulics remain powerful
What makes robot motion look natural: prediction, limited-horizon planning, and athletic maneuvers
Marc explains ‘natural’ movement as a combination of great hardware and forward-looking control, not just reactive servoing. He contrasts short-horizon planning for walking with longer-horizon planning needed for acrobatics like flips and sticking landings.
- •Natural movement requires prediction; purely reactive control is ‘backward-looking’
- •Limited-horizon planning (seconds) for locomotion: continuous replanning around obstacles and balance
- •Longer horizon for acrobatics: launch conditions must set up the entire maneuver
- •Historical arc: planar somersaults in the 1980s to more complex 3D maneuvers later
Biomechanics inspiration and robot morphology: knees, compliance, and passive dynamics
The conversation turns to how many joints/actuators robots need and what biology teaches—ostrich anatomy, cheetah running, and the role of compliance. Raibert discusses passive dynamics: letting the body participate in control, not forcing everything via computation.
- •Design trade-offs: simplify early to make progress, then add complexity as needed
- •Biology studies: horses, ostrich leg anatomy, cheetah speed/turning and tail dynamics
- •Knees introduced in BigDog; energy handling (negative work) and compliance considerations
- •Passive dynamics perspective: mechanics can ‘compute’ motion; the body participates in stability and efficiency
Boston Dynamics AI Institute: merging athletic intelligence with cognitive intelligence
Raibert lays out the AI Institute’s mission to fuse physical competence with higher-level cognition, so robots can learn tasks more like people do. Central themes include ‘watch-understand-do’ and reducing reliance on expert programmers through more general learning paradigms.
- •Two-part intelligence framing: athletic (real-time physical skill) + cognitive (planning, abstraction)
- •Robots today lack cognitive intelligence, increasing programming burden and limiting practicality
- •Moonshot direction: robots learning from human demonstration (on-the-job training)
- •Hardware still matters alongside learning and AI—embodiment is part of intelligence
Stepping-stones to moonshots: learning pipelines, uncertainty, and ML vs control
Marc explains how the institute balances ambitious goals with concrete yearly milestones. He describes task understanding via segmentation, starting domains (like bicycle repair), and argues robots must operate with incomplete specifications—then discusses how learning and model-based control may converge.
- •Methodology: ‘stepping-stones to moonshots’ to maintain feedback loops and momentum
- •Breaking demonstrations into actionable segments; mapping observations to reusable skills
- •Operating under uncertainty and lack of explicit environment/task models (navigation analogy)
- •Machine learning is promising, but physical robotics differs from language; best results still often come from MPC + emerging learned components
Building world-class robotics teams: fearlessness, diligence, intrepidness, and fun
Raibert defines the traits he looks for in teams and why robotics demands persistence through failure. He highlights BD’s testing ethos, raw video philosophy, and the practical necessities of building robust machines that can survive repeated crashes and iteration.
- •Technical fearlessness: take on problems you don’t yet know how to solve
- •Diligence: pursue robustness beyond narrow demos; test through perturbations and edge cases
- •Intrepidness: accept frequent failures and keep going; example—109 tries to nail an Atlas staircase demo
- •Technical fun and maker culture: hiring builders, not just credentialed specialists; ‘build it, break it, fix it’
Public demos, products, and the road ahead: Optimus, competition, cost curves, and life advice
Marc comments on Tesla’s Optimus and the broader humanoid ecosystem, stressing resources, ambition, and the emergence of real use cases. The episode closes with reflections on technology risk, his contrarian ‘Hawaiian shirt’ identity, and advice to aim for an unconstrained dream and work toward it.
- •Assessment of Optimus and other humanoid startups (Figure, Apptronik); importance of resources and ambition
- •Competition dynamics: category creation (e.g., quadrupeds) can benefit market adoption
- •Cost and reliability still have large upside; manufacturing expertise (e.g., Hyundai) may help
- •Risk vs opportunity framing for advanced AI/robots; career advice: start from your dream-without-constraints and get as close as possible
