Notes — Marc Raibert on Boston Dynamics, Legged Robots, and the Future of Robotics

Notes on Marc Raibert in conversation with Lex Fridman — Lex Fridman Podcast(https://lexfridman.com/marc-raibert/), 16 February 2024.

Four questions [Adler frame]

Q1 — What is it about as a whole? A career-spanning conversation with Marc Raibert — the roboticist who built the first hopping robots at Carnegie Mellon and MIT’s Leg Lab, then founded Boston Dynamics and, later, the Boston Dynamics AI Institute. It covers the technical history of legged robots (pogo-stick hoppers, BigDog, Spot, Atlas), the engineering philosophy behind them (dynamic balance rather than cautious static stability), the culture that produced them (technical fearlessness, diligence, intrepidness, fun), and where Raibert now thinks robotics needs to go: from athletic intelligence, which he judges largely solved, to the harder, largely unsolved cognitive intelligence a robot needs to work usefully and safely among people.

Q2 — How is it argued? Almost entirely through first-person narrative and worked technical example rather than abstract argument. Raibert reconstructs the physics of the pogo-stick robot’s three-part balance control (energy injection, foot placement, body-attitude correction) from memory, walks through the mechanical history from hydraulic BigDog to electric Spot, and answers Fridman’s prompts with specific anecdotes — the 109 attempts before Atlas cleared three steps, the DARPA funding pitch to Craig Fields, the Julia Child cooking-technique study Matt Mason ran on grasping. The throughline is a set of recurring maxims (‘you have to run before you can walk’, ‘why not?’, ‘technical fearlessness, diligence, intrepidness, fun’) that function as compressed engineering philosophy rather than formal claims.

Q3 — Is it true, in whole or part? As autobiography and engineering history it is Raibert’s own first-hand account and reads as reliable on dates, mechanisms, and institutional history — he was present for essentially all of it. Some numbers are recalled loosely and flagged as such by Raibert himself (‘I think it was’, ‘something like that’ recur around dates in the 1980s). His claims about the commercial and technical state of humanoid robotics — that cognitive intelligence, not hardware, is now the binding constraint, that Atlas outperforms Optimus — are informed expert opinion rather than independently verifiable fact, and are those of a competitor and board member, not a neutral observer [?]. The philosophical claim that superintelligent AI poses no special threat because ‘smarter’ does not imply ‘uncontrollable’ is asserted by analogy (to already-smarter humans, to cars) rather than argued through the alignment literature Fridman gestures at.

Q4 — What of it? The episode is the wiki’s first treatment of embodied robotics as opposed to software AI, and it supplies a durable distinction — Athletic Intelligence versus cognitive intelligence — that is likely to recur as humanoid-robotics episodes are added. It also documents, in unusual technical detail for a podcast, the actual control-theory mechanism (energy, foot placement, attitude) behind dynamic balance, which grounds later, vaguer references to ‘robots that can balance like animals’.

Glossary

Dynamic balance — keeping a legged robot upright by continuously predicting and correcting its motion in real time, as opposed to static or quasi-static stability, where a robot always keeps enough feet planted on the ground to be stable without any correction (like a table or chair). [§ Legged robots]

Athletic intelligence — Raibert’s term for the physical competence of a robot: mechanical design, real-time motor control, and energetics combined to move, balance, and manipulate objects. He contrasts it with cognitive intelligence — planning, perception, and judgement under uncertainty — and argues Boston Dynamics has largely solved the former while the latter remains the binding constraint on useful robots. [§ AI Institute]

Passive dynamics — a design idea (credited to Tad McGeer) in which a leg mechanism walks down an incline using only gravity, springs, and dampers, with no computer at all; the body itself, not just the control software, does part of the work of moving correctly. [§ Leg Lab]

Inverted pendulum (foot-placement control) — the calculation of where to place a hopping or running robot’s foot relative to its centre of mass, so that landing neither throws the body forward nor lets it fall backward — the same problem a pole-vaulter or high-jumper solves at takeoff. [§ Legged robots]

Stepping stones to moonshots — Raibert’s project-management principle at the AI Institute: never go more than about a year without a visible, if partial, result on the way to a multi-year target, so a team gets feedback and motivation along the way. [§ AI Institute]

Watch, Understand, Do — the AI Institute’s flagship research problem: a robot observes a human performing a task (starting with simple bicycle repair), segments the demonstration into discrete skills, and reproduces the task itself — robot on-the-job training, without an explicit model of the objects involved. [§ AI Institute]

Key claims by section

Early robots and the pull into legged locomotion [§ Early robots]

Raibert dates his conversion to robotics to 1974, seeing a disassembled robot arm in Berthold Horn’s MIT lab as a graduate student in brain and cognitive sciences; neurophysiology had felt insufficiently conceptual to him. He traces the origin of his life’s technical direction — legged robots — to a 1980s biomechanics conference where the only robotics talk showed a six-legged robot that kept at least three feet on the ground at all times, moving like ‘a table or a chair’. Raibert’s reaction was that this was nothing like how animals actually move: they bounce, fly, and predict their own motion to stay balanced, recycling energy through muscle and tendon rather than discarding it each step. That contrast — static, always-stable multi-legged walking versus dynamic, momentarily-unstable bouncing locomotion — became his founding research bet.

The pogo-stick robot and the mechanics of dynamic balance [§ Legged robots]

The first hopping robot (funded via Ivan Sutherland’s introduction to DARPA’s Craig Fields, who committed $250,000 on the spot in 1980) reduced legged locomotion to three coupled control problems, solved independently and combined: (1) how high to bounce, measured by estimating altitude at the top of the hop, since a pogo stick’s energy at that instant sits entirely in gravitational potential; (2) where to plant the foot relative to the body’s centre of mass on landing, analogous to a pole-vaulter’s or high-jumper’s takeoff calculation; and (3) how to correct the body’s tilt (attitude), which can only be done by applying torque between leg and body while the foot is on the ground — the physics simply do not allow attitude correction mid-air. Raibert is candid that the original solution was a rough sketch, not an optimised one, and that later researchers (including his own students) built faster, more capable variants on the same three-part structure.

From lab curiosity to BigDog: the practical engineering problem [§ BigDog / § Hydraulic actuation]

BigDog (funded by a DARPA biodynamics programme run by Alan Rudolph, with Martin Buehler recruited from McGill) was the first robot to bring power, computing, and hydraulic actuation fully onto the platform itself, rather than tethering it to an offboard hydraulic pump and VAX computer as the Leg Lab’s earlier quadruped had required. Raibert frames indoor operation (not damaging walls, not bumping furniture) as harder than outdoor trail-walking, where the only requirement is avoiding obstacles. Hydraulics, in his account, is an unfairly maligned technology — heavy innovation in valve design (much of it stalled since 1950s aviation) let Boston Dynamics build valves compact and efficient enough to power Atlas, in a package the size of a football producing five kilowatts.

From hydraulic to electric: Spot’s origin [§ Hydraulic actuation]

Spot’s design brief came directly from Larry Page (Google acquired Boston Dynamics in the LS3/BigDog era) asking for something around 60 pounds, electric rather than gasoline-hydraulic, and less intimidating in a domestic setting. This is presented as a single, simple conversational request rather than an extended product-development process.

Athletic intelligence versus cognitive intelligence [§ AI Institute / § Athletic intelligence]

Raibert’s organising claim for his post-Boston-Dynamics work: intelligence splits into an athletic part (mechanical design, real-time control, energetics — mastering movement and manipulation of the physical world) and a cognitive part (planning, perception, judgement under incomplete specification — the kind of everyday reasoning humans do constantly, like estimating a walk time and setting an alarm accordingly). He judges Boston Dynamics to have set the standard for athletic intelligence, and judges cognitive intelligence the industry’s actual bottleneck: robots remain ‘pretty dumb’, requiring extensive human programming for every task, which is why the AI Institute’s flagship project, Watch Understand Do, targets robots that learn tasks by observing humans rather than being explicitly programmed.

Passive dynamics and the role of mechanical design [§ Leg Lab]

Tad McGeer’s passive-dynamic walker — a legged mechanism that walks down a slight incline using only gravity, springs, and dampers, no computer at all — illustrates Raibert’s broader point that a machine’s mechanics, not only its control software, does real computational work: ‘the body is a participant in the motion’. He extends this to say that even computer-controlled robots perform better when the controller ‘knows what the mechanism wants to do’ rather than forcing motion against the hardware’s natural dynamics.

Team culture: fearlessness, diligence, intrepidness, fun [§ Building a team / § Videos / § Engineering]

Raibert names four traits of a strong robotics team: technical fearlessness (taking on problems with no known solution and finding an entry point); diligence (building a genuinely robust solution rather than one narrowly tuned to a demo — illustrated by Boston Dynamics videos of engineers deliberately tugging on ropes or pushing on doors to perturb a working robot); intrepidness (tolerating repeated failure — Atlas needed 109 attempts across six weeks to reliably clear three steps); and technical fun, which he treats as inseparable from good engineering. His video-editing philosophy (‘no explanation… do something worth showing and then show it’) reflects the same diligence-over-polish instinct: showing failure alongside success is what makes the success legible.

Machine learning versus traditional control [§ AI Institute]

Raibert acknowledges a post-ChatGPT surge of interest in learning-based approaches at the AI Institute (citing Marco Hutter’s Zurich reinforcement-learning group) but states that, as of the interview, the most impressive athletic performances shown publicly (Atlas’s dynamic manoeuvres) are still produced mostly by model-predictive control, a ‘traditional’ technique, not learned end-to-end — with learning increasingly blended in. [?] This is Raibert’s characterisation as of February 2024 and may not hold as a durable claim about the field.

On AI risk and superintelligence [§ Future of robotics]

Raibert’s stated position is that a smarter-than-human AI is not inherently threatening, drawing an analogy to already living among humans who are smarter than oneself without that being threatening, and to cars — a technology with large real harms (roughly 1.25 million deaths a year) that is nonetheless judged, on balance, a boon to humanity because the harms are being engineered down over time. He does not engage directly with the alignment-literature scenario Fridman poses (a system so much smarter that its actions become unpredictable before harm is visible) beyond restating that all technologies carry a risk/opportunity balance. [?]