In this podcast, Boston Dynamics founder Marc Raibert talks about the future of robotics: robots should move dynamically like animals, not slowly and safely. He's optimistic about robots moving from labs to real use. Key products mentioned: Spot (a four-legged robot, 1,500 sold, fully electric), Atlas (a humanoid that can backflip, took 109 tries to climb three steps), and Big Dog (an early milestone, carried 400 lbs for 20 miles). He disagrees that hardware innovation is over, stressing robots still need breakthroughs in perception and autonomous decision-making.
At a Glance Marc Raibert (founder and former CEO of Boston Dynamics, currently Executive Director of the Boston Dynamics AI Institute) discussed the future of robotics on the Lex Fridman podcast. His core argument is that by combining AI with advanced hardware, future robots will far surpass current
Marc Raibert (founder and former CEO of Boston Dynamics, currently Executive Director of the Boston Dynamics AI Institute) discussed the future of robotics on the Lex Fridman podcast. The core thesis: By combining AI with advanced hardware, future robots will far surpass current capabilities. He led the development of iconic legged robots such as Big Dog, Atlas, Spot, and Handle, emphasizing 40 years of accumulated research at CMU, MIT, and Boston Dynamics. Key conclusion: A new generation of robots must achieve breakthroughs in perception, motion control, and autonomous decision-making. The AI Institute will focus on cutting-edge research to push robots from the lab toward practical deployment.
Raibert argues that traditional robots are too "safe," and true breakthroughs require embracing dynamics and uncertainty.
He recalls deciding to enter the robotics field after seeing a disassembled Vikarm robotic arm at the MIT AI Lab in 1974. His early research was inspired by biological motion — after observing a hexapod robot moving "like a table" at a conference on biological legged locomotion, he concluded that approach was "wrong." Animal and human motion involves "bouncing and flying," requiring prediction and energy cycling.
This philosophy led to the "spring-loaded inverted pendulum" robot at CMU in the 1980s. Raibert collaborated with Ivan Sutherland, building a prototype with $3,000 in seed funding, followed by a $250,000 grant from DARPA. Key technical breakthroughs included: bounce energy management (controlling energy input via height estimation), foot placement calculation (analogous to a pole vaulter's body positioning), and posture control (applying torque only at ground contact).
Raibert extends this philosophy to manipulation: "If we look at robotic manipulation, 50 years have passed, and we are far from human-level performance... I think it's too safe." He cites an analysis by Matt Mason, director of the CMU Robotics Institute, of Julia Child's cooking videos — Julia used 40 different handling methods, none of which involved "grasping." "We need to try to break the safe framework of static grasping."
Raibert noted that Boston Dynamics was not a robotics company when founded in 1992, but rather a physics simulation firm. The real turning point came from its collaboration with Sony.
In the late 1990s, Boston Dynamics developed "AIBO Runner" for Sony's AIBO robot — a fast-running quadrupedal version. Although AIBO Runner itself "didn't look cool enough" (too small, requiring slow-motion playback to see the effect), it allowed the team to "get us back where we belonged."
Another key turning point was the abandonment of the surgical simulator business. The team developed a vascular anastomosis training device with force feedback, which was a huge success at a surgical trade show — "surgeons were almost lining up." However, the path to commercialization was blocked: surgeons believed they should charge for teaching, and the company (then with 12–15 people) lacked the resources to break into the hospital market. Raibert recalled: "We looked at each other and said, 'Let's build robots.'"
Big Dog was the true milestone for Boston Dynamics. DARPA's BioDynamics program received 42 proposals, but only three were funded. Raibert emphasized that the key figure was Martin Buehler (a professor at McGill University), who drove the testing culture of "getting out of the lab and into the mud" — "build, break, fix" became the company's creed. Big Dog was the first platform to integrate all systems: a gasoline engine, hydraulic system, and onboard computer, enabling autonomous balance and walking in the wild.
Raibert explicitly rejects the view that "hardware is no longer important": "Those who think hardware innovation is no longer needed are wrong."
He elaborated on Boston Dynamics' hydraulic technology expertise. Traditional hydraulic valve designs date back to the 1950s aviation sector, but Boston Dynamics' engineers redesigned the valve circuitry to make it smaller, lighter, and more efficient. The hydraulic power unit used in Atlas is only "the size of a soccer ball" (5 kg), yet delivers 5 kW of power, integrating a motor, pump, filter, heat exchanger, and valves.
The shift from hydraulic to electric stemmed from a question by Larry Page. During a meeting after Google's acquisition, Page asked: "Can you make a smaller, less intimidating version?" This led to Spot—Boston Dynamics' first fully electric, non-hydraulic robot. Raibert acknowledged his preference for hydraulics ("it's a great technology"), but electrification reduces maintenance difficulty and leakage risks.
On motion control, Raibert emphasizes "look-ahead" rather than "servo-based" approaches. Spot's control algorithm only looks 1–1.5 seconds ahead, but continuously iterates (recalculating every 0.1 seconds). For highly dynamic actions like backflips, momentum and rotation must be coordinated at takeoff, as there is almost no room for control mid-air. He recalled that the first planar robot backflip was achieved in 1985–86, while the 3D backflip was accomplished by Rob, a former gymnastics champion and MIT PhD student—"He had the physical ability and translated it into mathematics and algorithms."
Raibert divides robot intelligence into two parts: athletic intelligence and cognitive intelligence. Boston Dynamics has set the standard for athletic intelligence, and the AI Institute will focus on conquering cognitive intelligence.
Athletic intelligence encompasses mechanical design, real-time control, and energy management — "making robots move dynamically like animals." Cognitive intelligence, on the other hand, is the ability to "make plans, understand scenes, and learn from observation." Raibert identifies the core bottleneck of current robots: "Most robots are pretty dumb and require a lot of skilled people to program them."
The flagship project of the AI Institute is "Watch-Understand-Do": enabling robots to observe humans performing tasks, understand the sequence of actions, and then complete the tasks autonomously. The current starting point is "bicycle repair" — inserting the seat post into the frame and tightening it. Raibert believes this does not require an explicit model of the bicycle: "If I'm doing on-the-job training, I don't need to describe the room — we walk in, adjust the chair, and can navigate. Robots should be able to do the same."
Regarding methodology, Raibert adopts a "stepping stones to the moon" strategy: "If there is no visible status report (stepping stone) for more than a year, you need to simplify. Stepping stone results give you feedback and motivation." He acknowledges that doing machine learning on physical robots is different from language models — "pixel values aren't like words." However, the AI Institute has already established an office in Zurich, led by reinforcement learning expert Marco Hutter (Professor at ETH), to explore the fusion of learning and traditional control.
Raibert identifies four elements of an outstanding robotics team: technical fearlessness, diligence, resilience, and fun.
Raibert also shares his video-making philosophy: "Don't explain. If they can't see it, then it's not the right thing. Do something worth showing, then show it." He personally edited Boston Dynamics videos until 3–4 years ago, insisting on "no titles, no distractions."
| Position | Guest Stance | Key Data |
|---|---|---|
| Spot | Bullish (commercially viable, improved reliability) | 1,500 units deployed; fully electric; price still high but expected to decrease |
| Atlas | Bullish (benchmark for athletic intelligence) | 109 attempts to climb three steps; uses model predictive control; hydraulic power unit 5 kg / 5 kW |
| Big Dog | Historical achievement | Gasoline engine; can carry 400 lbs (actual test reached 1,000 lbs); range of 20 miles |
| LS3 | Historical achievement | Can carry 400 lbs; gasoline-powered |
| Wildcat | Historical achievement | Quadruped robot top speed 19 mph |
| Optimus (Tesla) | Neutral (respects Elon, but currently lags behind Atlas) | No specific data disclosed |
| Figure | Neutral (good team) | Not disclosed |
| Apptronik | Neutral (good team) | Not disclosed |
| Anymal (ANYbotics) | Neutral (already purchased by AI research institute) | Not disclosed |
| Sony AIBO | Historical collaboration | AIBO Runner project; weekly meetings with Japanese team via ISDN line (10 Hz frame rate) |
1. Raibert: Robot motion requires "look-ahead" rather than "servo" control — Spot's control system only looks 1–1.5 seconds ahead, but recalculates every 0.1 seconds, continuously iterating. Servo control "looks backward," while dynamic motion requires "predicting future states."
2. Raibert: Hardware innovation is far from over — "Anyone who thinks no more hardware innovation is needed is wrong." Boston Dynamics redesigned the hydraulic valve (unchanged since the 1950s), achieving 5 kW power output from a 5 kg unit.
3. Raibert: The golden rule for robot videos is "no explanation" — "If they can't see it, then it's not the right thing. Do things worth showing, then show them." He personally edited Boston Dynamics videos until 3–4 years ago, insisting on no titles and no distractions.
4. Raibert: It took Atlas 109 attempts and 6 weeks to climb three steps — "We filmed every single attempt — including the robot collapsing right after startup, with oil splattering." This underscores the importance of "intrepidness."
5. Raibert: Robot intelligence is divided into "motion intelligence" and "cognitive intelligence" — Boston Dynamics has set the standard for motion intelligence, while the AI Institute will tackle cognitive intelligence (observe-understand-execute). The current bottleneck is that "most robots are quite dumb."
6. Raibert: The AI Institute adopts a "stepping stones to the moon" strategy — "If there's no visible status report for more than a year, you need to simplify. Stepping stone results give you feedback and momentum."
7. Raibert: He is skeptical about the threat of AGI — "There are people in the world smarter than you. Do you feel threatened?" He believes technology always carries both risks and opportunities, drawing an analogy to cars (1.25 million deaths per year) which remain a net benefit to humanity.
8. Raibert: Engineers are "paid four times" — the satisfaction of craftsmanship, the fulfillment of impacting the world, the joy of teamwork, and the salary. "What could be better than that?"