Marc Raibert: Boston Dynamics and the Future of Robotics | Lex Fridman Podcast #412

Marc Raibert recounts how Boston Dynamics turned dynamic locomotion into practical robots.

Core Philosophy

• Raibert rejected cautious, statically stable machines after watching slow six-legged walkers and chose bouncing, flying gaits instead (0:00-10:03).
• The motto "you have to run before you walk" proved that aggressive motion can be easier to stabilize than slow walking (40:19-50:20).

Key Robots and Milestones

• One-legged pogo-stick prototypes secured DARPA funding and led to BigDog, LS3 and the electric Spot (10:07-30:13).
• Atlas demonstrated robust dynamic climbing after 109 falls, proving hardware durability enables rapid iteration (1:10:29-1:20:30).

Future Direction

• The new Boston Dynamics AI Institute pairs athletic intelligence (mechanics and control) with cognitive intelligence so robots can learn tasks by watching humans (50:22-1:00:23).

00:00:00Major Topic: Movement, Hardware, and the Evolution of Boston Dynamics

00:00:00 The Aesthetic of Movement and the Hardware Imperative

• The opening premise contrasts BigDog/LS3 with the beauty and naturalness of movement in Boston Dynamics’ robots. The question is framed: what does it take to make a machine look natural when walking, running, or flipping?
• Marc Raibert responds that hardware innovation is essential: some people think hardware is obsolete, but he argues that both mechanical design and control must progress in tandem to achieve lifelike motion.
• The early remark from Lex highlights the capability: LS3 can carry around 400 pounds, yet there are anecdotes of carrying far more (e.g., a test showing 1,000 pounds). This signals the engineering margin and the showmanship of reliability testing.

> "So BigDog became LS3, which is the big load-carrying one." (00:00:00) > "Just a quick pause. It can carry 400 pounds." (00:00:00)

00:10:00 Major Topic: Marc Raibert’s Background and the Lab's Philosophy

• The host introduces Marc Raibert as the founder/CEO of Boston Dynamics, with a long history at CMU, MIT Leg Lab, and MIT/Carnegie contexts. The framing emphasizes a career-long focus on legged locomotion and the pursuit of future generations of robots.
• Early roots are explored: Raibert describes growing up around tools (his father’s tinkering), his first spark with robotics in 1974 when he saw a vyke arm disassembled, and how that image ignited his lifelong pursuit. The year 1974 is explicitly recalled as a turning point.

> "I followed him back to his lab. And on the table was a vyke arm, robot arm taken apart in probably 1,000 pieces. And when I saw that, from that day on, I was a roboticist." (00:10:00) > "It was really during graduate school when I followed a professor back from class... And when I saw the arm, I just thought, you know, this is it." (00:10:00)

00:20:00 Major Topic: The Tension Between Cognitive Science and Robotics Approaches

• Raibert notes the tension between the brain-and-cognition (BCS) approach and the robotics approach. BCS (brain and cognitive science) seeks to understand intelligence biologically, while robotics pursues intelligent behavior in machines. He mentions bridges like the Center for Brains, Minds and Machines (CBMM) and David Maher’s models as attempts to bridge the gap.
• The conversation acknowledges how early rails of AI and robotics intersect, and how the bridge fluxed and waned over decades, with renewed strength in the last five years.
• A side reflection on early tinkering shows how Raibert balanced exploration (function) with aesthetics, noting that sometimes the machines looked like junk until movement lent them life. The takeaway: aesthetics can compound lifelike perception once functional motion exists.

> "BCS is now morphed a bit... the bridge between those two groups." (00:20:00) > "The heart of it is, you know, making them do things that are interesting." (00:20:00)

00:25:00 Major Topic: Early Personal Tendencies and Notions of Function vs Form

• Raibert discusses early projects and the balance between function and aesthetics. He recalls a childhood penchant for gadgets, but recognizes that the real drive is mobility, dexterity, perception, and intelligence. He notes a shift: early on, function dominated, but later he learned that the aesthetics of the machine can enhance lifelike perception when movement is present.
• A key theme: robots should do things that are interesting, but there’s a balance between what’s visually appealing and what’s technically required.

> "There was a time though, when I was... I thought that function was everything. Mobility, dexterity, perception... and nothing else mattered." (00:25:00)

00:30:00 Major Topic: The Early Leg Lab and the Inception of Hopping Robots

• Raibert recounts the founding timeline: Leg Lab originated at CMU, with the first hopping machine around 1982, a 3D Quadruped around 1983, and a more mature version in the mid-80s. The timeline underscores a long arc of development, deep trial-and-error, and iterative improvements.
• The pivotal turn from CMU to MIT/JPL to the founding of Boston Dynamics is described, including collaboration with Ivan Sutherland, and a risk-taking moment with DARPA seed funding.
• The pogo-stick robot is presented as the initial bold step; a simple design evolved into a more capable device. The initial funding (DARPA) came after presenting a skeletal model to Craig Fields, who subsequently funded the project.

> "What was on the technical side... I hired Ben Brown. We came up with a thing that could start to work... It wasn't that hard to make it balanced once you get the physical machine to be working well enough." (00:30:00) > "The first talk I attended showed a six-legged robot that walked very slowly... I said, 'wow, that's wrong'... We should bounce and use the springiness in our legs." (00:30:00)

00:40:00 Major Topic: Core Control Problem and 3D Balancing for Pogo-Stick Robots

• The three-headed control challenge of a pogo-stick leg: (1) bouncing energy storage in spring, altitude, and velocity; (2) foot placement relative to the center of mass to achieve the correct trajectory; (3) maintaining upright attitude to prevent uncontrolled rotation.
• The concept of balancing arises as a dynamic control problem in 3D; early versions tackled the 2D problem, then progressed to 3D with a sense of the three critical components. The top-of-hop energy is in height, enabling energy planning; foot placement is akin to guiding a pole vaulter/ high jumper to clear the obstacle; attitude control keeps the structure from rotating out of control.
• Raibert notes limitations and acknowledges later developments: more advanced algorithms were developed by students and others to push speed and obstacle-climbing capabilities.

> "There are three things going on: bounce energy, foot placement, and attitude control. If the foot hits wrong, you're off; organize the energy so you go mostly up; keep the body upright so you don't rotate away." (00:40:00) > "The key thing the pole vaulter has to do is get its body to the right place when the pole gets stuck... to go mostly up." (00:40:00)

00:50:00 Major Topic: From Planar to 3D, and the Early Doubts of Humanoid Robots

• The team’s experiments included planar (2D) hopping devices before extending to 3D. The early planar work provided the groundwork, and the 3D extension was nontrivial but achievable. Key figures include Roy Rob (a gymnast-turned-engineer) who helped translate human movement intuition into robotic control, including 3D somersaults and stabilization strategies.
• The discussion touches on the broader challenge in robot manipulation: after five decades, human-level manipulation remains elusive; the problem domain benefits from risk-taking and purposeful deviations from overly safe approaches.
• Raibert reflects on the early stance toward humanoids: skeptical about form vs function; the decision to pursue humanoids came later, when social impact and interaction opportunities became clearer (e.g., public interest in humanoids and interaction with bystanders).

> "Our initial thinking on humanoids wasn't enthusiastic because the form wasn't as important as functionality." (00:50:00) > "The connection and impact with other people... there's a special thing that goes on; even non-technical people feel a particular pull toward humanoid robotics." (00:50:00)

01:00:00 Major Topic: BigDog, LS3, and the Evolution of Powertrains

• BigDog is highlighted as the catalyst that put Boston Dynamics on the map, propelled by DARPA’s biodinautics program via Alan Rudolph. Martin Buehler’s arrival helped move BigDog from the lab to real-world mud—an essential step to accept risk, test under real conditions, and iterate quickly.
• The BigDog platform was originally hydraulic-powered with the engine driving a hydraulic pump; the control system integrated the power, hydraulics, and computation on-board. The shift to a battery-powered, all-electric variant (the next generation) began the move toward home-friendly size and applications.
• The LS3 variant scaled up payloads dramatically (designed for 400 pounds; tested with up to 1,000 pounds). The vehicle’s rugged real-world testing was performed on challenging trails (e.g., Guadalcanal Trail at Quantico).

> "BigDog was the first thing that worked. We brought it out of the lab into the mud; we tested on the Guadalcanal Trail to meet the DARPA milestone." (01:00:00) > "BigDog had hydraulics; LS3 transitioned to electric power, scaling down the size and making it more house-friendly." (01:00:00)

01:10:00 Major Topic: The Switch to Electric and the Birth of Spot

• The transition from BigDog/LS3 to Spot is discussed as a major inflection point: reducing weight (roughly 60 pounds) to make a household-friendly robot, moving away from hydraulic actuation to electrics, while maintaining efficient actuation and control.
• The role of Larry Page in pushing toward a smaller, more approachable robot is highlighted: he asked if a smaller, non-intimidating robot could be made for a home. The answer: yes, and development followed.
• The engineering challenge of replacing hydraulics with electric power (while still delivering sufficient power and precision) is framed as an opportunity to rethink actuation architecture. The newly designed compact hydraulic power package (a small unit the size of a football, ~5 kg, 5 kW) demonstrates the ongoing hardware innovation and integration.

> "Larry asked if you could make something that weighed 60 pounds that would not be so intimidating if you had it in a house; we could do that." (01:10:00) > "We loved hydraulics, but moving to electric power required new valve technology and an integrated power unit; it opened up new design space." (01:10:00)

01:20:00 Major Topic: Natural Movement, Predictive Dynamics, and the Horizon of Planning

• The discussion emphasizes the role of prediction and forward-looking planning in making motion look natural. Raibert calls out the importance of a short planning horizon (a couple of seconds) for Spot, with continuous re-planning as the robot senses its environment. He contrasts this with longer-horizon tasks like somersaults, which require careful momentum and rotation planning at launch.
• The concept of dynamic, predictive control contrasts with pure servoing (reactive control). The emphasis is to anticipate future states and correct for dynamic interactions that will happen in the next moments, rather than only responding to current errors.
• A reference to Tad McGeer’s passive-dynamics work is noted: a mechanical system can implement some walking with dampers and springs, acting as a mechanical computer, reducing computational load and increasing efficiency. Raibert emphasizes the value of co-design between mechanical dynamics and control strategies.

> "The horizon for Spot is only a couple of seconds; we look ahead a little, then re-evaluate as conditions change." (01:20:00) > "A mechanical system can walk down an incline with springs and dampers—it's a mechanical computer; that's the essence of passive dynamics." (01:20:00)

01:30:00 Major Topic: Atlas, Wildcat, and the Path Toward Humanlike Locomotion

• The narrative covers Atlas, the ongoing pursuit of more humanoid-capable robots, and the balance between function and aesthetics. Raibert notes initial skepticism toward humanoids due to cosmetics, but later acknowledges the powerful social impact and the demonstration value of humanoid platforms.
• The Wildcat project is recalled as a fast quadruped (~19 mph) that pushed the envelope on speed, though it was loud due to a racing go-kart engine.
• The role of joints and knees is explored: the early Leg Lab designs didn’t emphasize full knee architecture; later, BigDog introduced kneelike joints, while Spot kept a simpler, more energy-driven approach. The concept of passive dynamics and energy storage returns as a theme for efficient movement.

> "I wasn't an enthusiast for humanoids at first because the form wasn't as important as the functionality." (01:30:00) > "BigDog to Spot involved moving from hydraulic power to electric to reduce weight and enable home-scale use." (01:30:00)

01:40:00 Major Topic: The AI Institute and the Two-Part View of Intelligence

• A major pivot: Raibert discusses the Boston Dynamics AI Institute and the vision to fuse athletic (physical) intelligence with cognitive (planning, understanding) intelligence. He defines intelligence as having two parts: the athletic, physical execution, and the cognitive, planning and understanding. The Institute aims to teach robots to watch humans perform tasks, understand what is happening, and replicate the task; a form of on-the-job training for robots.
• The discussion underscores the need for robots that can operate effectively under uncertainty and without heavy human-slaved specification; the Institute’s research aims to enable more autonomous learning and adaptation.

> "Intelligence has two parts: athletic and cognitive. Boston Dynamics is the standard for athletic intelligence; the AI Institute aims to combine that athleticity with cognitive understanding." (01:40:00)

01:50:00 Major Topic: Machine Learning, Reinforcement Learning, and the Zurich Office

• The role of machine learning in robotics is discussed. While there’s growing optimism post-ChatGPT, Raibert emphasizes that physical robotics poses different challenges than language tasks. He notes ongoing ML efforts, including Marco Hutter at ETH and the Zurich office, as part of a balanced approach (learning plus traditional control like MPC).
• The institute is positioned to blend traditional control methods with learning-based strategies to achieve robust, reliable, and efficient behavior in real robots.

> "There’s a lot of interest in machine learning approaches, but physical robot control often relies on traditional controls like model predictive control; we’re aiming for a mating of the two." (01:50:00)

02:00:00 Major Topic: Building Teams and Four Pillars of Great Engineers

• Raibert highlights four essential traits for a great robotics team: technical fearlessness, diligence, intrepidness, and fun. Each is explained with practical meaning:

- Technical fearlessness: willingness to tackle problems without a clear entry path, using stepping stones to front-load risk and learning. - Diligence: insistence on robust, generalizable solutions rather than brittle, narrow ones; the team’s testing ethos is exemplified by Spot’s door-opening and rope-tugging failures. - Intrepidness: a willingness to press forward despite setbacks, including long trial Series (e.g., Atlas’s 109 attempts to complete difficult steps) before achieving reliable success. - Fun (technical fun): enjoying the craft of engineering and the satisfaction from tangible impact; engineering as a craft that yields personal fulfillment and organizational success.

> "Technical fearlessness means being willing to take on a problem you don’t know how to solve, and then learning from stepping stones to eventually solve it." (02:00:00) > "Intrepidness: you have to stick to it and keep trying; it’s the most important ingredient." (02:00:00) > "Get paid to have fun" (01:10:29)

02:30:00 Major Topic: Demonstrations, Video-First Communication, and Authentic Progress

• The videos from Boston Dynamics are treated as a central communication channel. Raibert emphasizes showing real progress, including failures, to illustrate robustness and the learning curve. He notes the philosophy of “no explanations” in video—let the viewer see what the robot can do and how it responds under perturbation. The honesty of producing multiple takes and capturing both failures and successes in videos is a central value.
• He discusses the balance between public perception and engineering reality, acknowledging how videos shape sentiment and expectations, and how the team addresses criticisms around BigDog videos where a human operator was visible.

> "The video philosophy is: if it’s worth showing, show it without over-explanation; if the robot is doing something worth watching, let viewers see it." (02:30:00) > "We shoot all the tests; Atlas climbing steps took 109 tries; you see the evolution, not just the best moment." (02:30:00)

03:00:00 Major Topic: The Social and Cultural Meaning of Robotics; Competition; and the Future of Home Robots

• Raibert explores broader questions about the social acceptability and commercial viability of humanoid and legged robots. He points to the challenge of achieving real use cases that are money-making in robotics beyond industrial arms and warehouses: home-care, assistance, and companionship (e.g., Spot-like platforms in the home).
• He stresses the importance of the category developing through real use cases, cost reductions, and reliability improvements—along with societal acceptance through clear boundaries and privacy norms. Hyundai’s ownership of Boston Dynamics is noted as a factor in manufacturing scale and cost reduction.
• The discussion returns to the open question of AGI (artificial general intelligence). Raibert emphasizes that intelligence is multi-faceted and that even if machines become smarter in certain domains, alignment and safety remain critical concerns. He likens the challenge to real-world risk/benefit tradeoffs (e.g., the relative harms of cars today versus their societal benefits).

> "In the quadruped space, competition is a good thing; it clarifies use cases and drives improvements; the market will decide which quadruped best fits each need." (03:00:00) > "We should not fear AGI as an immediate existential threat; instead, we should balance opportunities with risk and solve practical safety problems as we go." (03:00:00)

03:30:00 Major Topic: The Hawaiian Shirt, Contrarian Culture, and Advice for Young Engineers

• The Hawaiian shirt is explained as a personal symbol of contrarianism and independence—an attitude Raibert embraces to challenge norms and stay curious.
• He provides practical career advice to young people: if resources and constraints were removed, what would you do? Start there and then see how close you can come to realistic execution. The message emphasizes broad opportunity and the importance of pursuing meaningful, long-term work over chasing short-term status.

> "The Hawaiian shirt is a symbol of why not." (01:40:45) > "If you had no constraints on you, what could you imagine doing? Start there and see how close you can get to what's realistic." (01:40:45)

04:00:00 Major Topic: Closing Thoughts and the Vision for the Future

• The interview closes with reflections on the future of robotics, potential collaborations between Atlas and Optimus, and the ongoing mission of the AI Institute. Raibert expresses openness to cross-pollination between platforms and emphasizes the importance of demonstration, storytelling, and shared events like robot meetups.
• The final sentiment touches on the meaning and joy of engineering: making metal come alive, the beauty of elegant motion, and the reward of building something that others admire and rely on.

> "Bringing metal to life or a machine to life is magical; the dancing videos show progress and public engagement, and we’ll do more." (04:00:00) > "I’d love to host Atlas and Optimus meetups; the collaboration between platforms could accelerate the field." (04:00:00)

04:15:00 Major Topic: Resources Mentioned and People/Institutions Involved

• The conversation traverses through a broad ecosystem: JPL, CMU, MIT, MIT Leg Lab, CMU Leg Lab, DARPA programs (BigDog/LS3), Sony and Aibo, LittleDog, RiSE, Atlas, Spot, Handle, ANYmal, and the new Boston Dynamics AI Institute.
• The collaborative network includes Ivan Sutherland, Craig Fields (DARPA PM), Alan Rudolph (DARPA biodinautics program), Martin Buehler (CMU/Montreal, BigDog program), Roy Rob (MIT/ROM gymnast-turned-robotics student), Tad McGeer (passive dynamics), Marco Hutter (ETH Zurich), and Zurich office leadership. The corporate backdrop includes Google’s acquisition of Boston Dynamics and Hyundai’s current ownership.
• The discussion references the broader ecosystem: CBMM (Center for Brains Minds and Machines), Sensible Technologies (force-feedback devices), QRIO and Aibo (Sony), BigDog/LS3, LittleDog, Spot, Atlas, Wildcat, and the role of YouTube videos in public perception.

> "There are 1,500 of them out there now" (01:00:24) > "The Center for Brains, Minds and Machines (CBMM) is bridging neuroscience and AI." (01:00:24) > "We opened an office in Zurich; Marco Hutter leads reinforcement learning for robots there." (01:50:00)

04:30:00 Major Topic: Summary of Core Insights and Takeaways

• Hardware still matters: advanced actuation, energy storage, and power management are as critical as software for achieving lifelike motion.
• Dynamic, predictive control yields more lifelike movement than pure servoing; future robots must plan a few seconds ahead and replan as conditions change.
• The path toward humanoid and quadruped robotics involves a careful balance of risk, robustness, and social relevance; early demonstrations require transparency about failures to teach the public and stakeholders.
• The AI Institute’s guiding principle is to fuse athletic (motion) intelligence with cognitive learning and perception, enabling robots to watch humans perform tasks and then replicate them via on-the-job learning.
• Team culture matters: fearlessness, diligence, intrepidness, and fun are foundational to a high-performing robotics team.
• The long arc of innovation includes a willingness to fail, to iterate across platforms, and to maintain a public-facing narrative that emphasizes genuine capability rather than perfected perfection.

> "Athletic intelligence is about the body and the mechanics; cognitive intelligence is about understanding and planning; the AI Institute aims to merge both." (01:40:00) > "There is beauty in movement when the robot behaves as if it has life; we’re still far from human motion, but we’re getting closer in controlled dimensions." (00:00:00)

04:45:00 Major Topic: Final Reflections on the Meaning of Robotics

• The final reflections touch on meaning and fulfillment in engineering: fun, craft, social impact, and the wonder at making machines that can emulate life. Raibert reiterates the sense of purpose in creating the next generation of robots and invites the audience to imagine what it would look like for robots to learn by watching humans perform tasks in real-world settings.

> "Engineering is a craft that can be deeply satisfying; the joy comes from making something that others admire and that can impact the world." (04:45:00) > "The ride is short; might as well have fun." (04:45:00)

05:00:00 Resources Mentioned

• BigDog, LS3, Spot, Atlas, LittleDog, Aibo, Aibo runner, RiSE, Handle, Wildcat, ANYmal, and other Boston Dynamics platforms.
• DARPA programs and funding: biodinautics (BigDog/LS3), DARPA PM Craig Fields, Quantico and Guadalcanal Trail trials.
• Hardware and software ecosystems: hydraulic actuation, electric actuation, new valve designs, integrated power packs, and model-predictive control (MPC) in combination with learning.
• External institutions and companies: MIT Leg Lab, CMU Leg Lab, MIT AI Lab, JPL, Sony (Aibo), Sensible Technologies (force feedback devices), CBMM, ETH Zurich (Marco Hutter), Hyundai (current owner of Boston Dynamics), Google (historical owner during Google's period of involvement).
• Educational and collaborative initiatives: Choreo Robotics (Brown University), Sydney Skybetter, Stephanie Tellex, and the broader robotics education ecosystem.

> "We opened an office in Zurich; Marco Hutter leads reinforcement learning for robots there." (01:50:00) > "There are 1,500 of them out there now" (01:00:24)

05:30:00 Notable Quotes (Selected VERBATIM)

• > You have to run before you can walk. (00:40:19) — Speaker: Marc Raibert
• > There are 1,500 of them out there now. (01:00:24) — Speaker: Marc Raibert
• > The Hawaiian shirt is a symbol of why not. (01:40:45) — Speaker: Marc Raibert
• > Get paid to have fun. (01:10:29) — Speaker: Marc Raibert
• > Intelligent has two parts: athletic and cognitive. (01:00:24) — Speaker: Marc Raibert

Resources Mentioned

• DARPA biodinautics program (BigDog/LS3), Craig Fields (DARPA PM), Alan Rudolph (DARPA program lead), Martin Buehler (CMU/Montreal)
• MIT Leg Lab, MIT AI Lab, CMU Leg Lab, JPL, Sony (Aibo, Aibo runner), Five additional robotics platforms: BigDog, LS3, LittleDog, Spot, Atlas, Wildcat, RiSE, Handle, ANYmal
• Sensible Technologies (force-feedback devices), Center for Brains Minds and Machines (CBMM), ETH Zurich (Marco Hutter), Choreo Robotics (Brown University), Sydney Skybetter, Stephanie Tellex
• External resources and brands: Google (formerly), Hyundai (current owner), RC/Electronic components, QRIO (Sony), Aibo (Sony)

Key Takeaways (Punchy bullets)

• Hardware innovation remains crucial for lifelike robot movement; software alone cannot compensate for poor actuators or power systems.
• Dynamic, short-horizon planning plus continuous re-planning is essential for real-world locomotion with obstacles.
• The pogo-stick lineage laid the groundwork for modern legged robots and taught the importance of balancing, foot placement, and body attitude.
• There is a deliberate shift from hydraulic to electric actuation to enable smaller, more home-friendly robots without sacrificing performance.
• The AI Institute’s two-part view of intelligence (athletic and cognitive) frames a holistic approach to future robots: fast physical action plus understanding and learning from humans.
• Incremental, stepping-stone progress is preferred over moonshots alone; tangible milestones and demonstrable progress sustain momentum.
• Robust, reliable robotics demands serious diligence and many iterations (Atlas’s 109 trials to success; many failures before breakthroughs).
• Public demonstrations, even when imperfect, can catalyze broader interest and real-world adoption—transparency about failures is a feature, not a bug.
• The future of home robotics rests on cost reductions, reliability, and clear business cases; competition in quadrupeds/humanoids will drive cost discipline and feature breadth.
• The meaning of robotics blends craft, innovation, and social impact: humans live in a world that robots can augment, extend, and learn from by watching and imitating.

Action Items (Concrete, do-this steps)

• Build a compact timeline tracing BigDog → LS3 → Spot → Atlas, and annotate with payloads, powertrain changes, and control strategies.
• Create a diagram of the pogo-stick control loop: energy storage (spring/altitude/velocity), foot placement planning, and attitude control; annotate how each component affects stability.
• Compare hydraulic vs electric actuation: enumerate tradeoffs (weight, efficiency, control bandwidth, packaging, and reliability) and map to Spot’s design decisions.
• Summarize Raibert’s athletic vs cognitive intelligence framework; draft a one-page explainer you could share with peers.
• Compile a list of “stepping stones” projects from the AI Institute plan; for each, outline a measurable milestone and a rough timeline.
• Draft discussion prompts exploring the social implications of robotics in the home (privacy, safety, adoption barriers) to accompany demonstrations.
• Build a glossary of the key terms and acronyms (see Glossary below) to standardize vocabulary for study.

Mental Models

• { name: "Athletic vs Cognitive Intelligence", detail: "Two-part model: athletic (physical capabilities, dynamic control, energy management) + cognitive (perception, planning, learning). The institute’s goal is to fuse both for robust, adaptable robots." }
• { name: "Stepping Stones to Moonshots", detail: "Progress in long-term goals should be broken into tangible milestones with periodic demonstrations and status reports; avoid waiting for a monolithic breakthrough before sharing progress." }
• { name: "Passive Dynamics", detail: "A mechanical design philosophy where the body’s natural physics (springs, dampers, inertia) contribute to motion, reducing active control requirements and improving efficiency." }
• { name: "Limited-Horizon Planning", detail: "Plan a short horizon (a few seconds) and re-plan as new information arrives; balance planning depth with real-time sensing and obstacle handling." }
• { name: "Model-Based + Learning Hybrid", detail: "Blend traditional control methods (MPC) with learning-based adaptation to capture both reliability and adaptability in real environments." }

Glossary

• BigDog: A large quadruped robot designed for rough-terrain locomotion; powered by an on-board gasoline engine driving a hydraulic power system.
• LS3: The Load-carrying version of BigDog, designed to carry heavy payloads (up to 400 pounds planned; demonstrated with far more).
• Spot: A smaller, all-electric quadruped designed for household-scale applications, with on-board power and computation.
• Atlas: A humanoid robot platform built by Boston Dynamics, focusing on human-like locomotion and manipulation.
• LittleDog: An early smaller dog-like robot from Boston Dynamics (linked to the Aibo lineage).
• Aibo runner: Sony’s Aibo robot with components repurposed for robotic experimentation; an early link to Boston Dynamics’ hardware development.
• RiSE: A climbing robot project funded by DARPA as part of biodinautics; demonstrates advanced terrain interaction.
• ANYmal: A legged robot platform from a separate company (Marco Hutter’s team) used in collaboration with Boston Dynamics.
• MPC (Model Predictive Control): A planning/control method that optimizes future control actions given a model of the system dynamics.
• OJT (On-the-Job Training): A learning paradigm whereby a robot learns tasks by observing humans perform them and then replicating the actions.
• CBMM (Center for Brains Minds and Machines): A center at MIT focusing on bridging neuroscience and AI; a conceptual bridge referenced in the talk.
• Passive Dynamics: A design principle where the physical structure of a robot contributes to motion with minimal active control, leveraging natural dynamics.
• Passive dynamics + actuation: The combination of mechanical design that harnesses energy storage and dissipation with active control for robust locomotion.

Quotes (verbatim excerpts with timestamps)

Topics

• Robotics
• Legged Robots
• Autonomous Systems
• AI Institute
• Humanoid Robots
• Machine Learning in Robotics
• Dynamic Locomotion
• Hardware-Software Co-Design
• Sensing and Perception in Robots
• Human-Robot Interaction