Playter, Robert
▶ Watch on YouTubeCore ideas
Playter traces his fascination with robotics to a formative visit to Marc Raibert's MIT Leg Lab, where seeing a robot perform a somersault connected instantly with his own background in gymnastics. He frames the core insight of legged robotics — one he attributes directly to Raibert, starting with a simple bouncing pogo-stick machine — as building machines whose physics naturally "want" to do the right thing rather than fighting the physics with brute-force control, and argues this same principle explains why Boston Dynamics' robots look organic and lifelike: they work with dynamic instability (allowing a fall and catching it with the next foot placement) rather than rigidly preventing any tipping, which he says is also, incidentally, what makes the motion more efficient and robust. He identifies three cultural values inherited from Raibert's lab that persist at the company today: pursue problems out of genuine curiosity regardless of near-term commercial payoff; find the simplest possible version of a hard problem (the pogo stick before the galloping horse); and be willing to break expensive hardware repeatedly as part of learning, rather than treating robots too preciously to experiment with.
He describes the surprising multi-decade difficulty of achieving natural-looking humanoid walking: even after building the early "Petman" prototype (2008) with deliberate heel-strike gait mechanics, Boston Dynamics didn't achieve motion that looked truly natural until roughly 2022, when a new walking-control approach on the latest Atlas produced natural gait as a side effect of a different underlying methodology rather than as a directly hand-tuned goal. He explains several concrete technical challenges: humanoids are "underactuated" (unlike a hovering drone, forces can only be applied through foot contact with the ground, not in arbitrary directions at will); a fully extended leg approaches a mathematical singularity that complicates control; and picking up heavy or irregular objects (a toolbag, a barbell) requires explicitly modeling the shifted center of mass and inertia and planning several seconds ahead to preserve balance — a problem absent in Boston Dynamics' earlier quadrupeds, where a rigid, heavy body and comparatively light legs decoupled leg motion from body dynamics. He credits recent gains in acrobatic behaviors (backflips, the "sick trick" front-flip-with-a-half-twist) to model predictive control, which lets the robot continuously replan its trajectory in real time — including mid-air, where momentum is fixed but the robot can still redistribute limb motion to control its landing configuration — replacing the earlier, much slower process of manually iterating trajectories by hand.
Playter walks through Boston Dynamics' history: founded in 1992 as a physics-based simulation company (not a robotics company) for roughly a decade; the 2003 DARPA contract that produced BigDog, the company's first fully self-contained legged robot (onboard power, computing, and hydraulics), whose early versions used loud, unmuffled go-kart engines chosen purely to save weight; a subsequent DARPA-funded portfolio of quadrupeds (LS3 for heavy load-carrying, Cheetah for speed) that let the team derive general scaling principles for legged design; the DARPA Robotics Challenge (2015), which Playter describes as humbling because ordinary human tasks (like getting in and out of a car) proved extremely hard for a general-purpose humanoid, illustrating the fundamental tension between general-purpose versatility and task-specific optimization; and the sale to Google, where Larry Page's "toothbrush test" push for an inexpensive consumer product led the company to instead pursue Spot as an industrial platform, since the underlying research philosophy (build the best possible machine to surface real problems, rather than cutting corners on hardware quality) was incompatible with hitting a low consumer price point quickly. He also recounts the company's abandoned bipedal-wheeled warehouse robot, Handle — an elegant machine using momentum-control mathematics derived from Atlas balance research, ultimately shelved as commercially too slow — in favor of Stretch, a non-balancing wheeled mobile-manipulator purpose-built for unloading shipping containers, which reached customer commitments (DHL, Maersk, Gap, and others) for hundreds of units essentially as soon as it was described, unlike Spot's longer search for a killer application (ultimately settling on industrial inspection — thermal imaging, acoustic leak detection, gauge reading — for customers like AB InBev and GlobalFoundries).
On the business of robotics, Playter is direct that most robotics startups fail because they can't reach the sales volume (thousands to tens of thousands of units) needed to sustain a hardware business, and that Boston Dynamics' current owner, Hyundai (with SoftBank retaining a 20% stake), is explicitly betting on a longer-term vision rather than demanding immediate profitability — a balance Playter describes internally as pursuing "self-determination" (staying profitable enough to keep steering the company's own direction). He frames Elon Musk's entry into humanoid robotics (Tesla Optimus) as broadly validating and energizing for the field, noting competing startups now advertise ex-Boston-Dynamics staff as a credibility signal, and sees Tesla's manufacturing-scale expertise as a genuine complementary strength while arguing Boston Dynamics is similarly backed by an automotive manufacturer (Hyundai). On AI, he distinguishes the company's use of reinforcement learning (already applied to locomotion, manipulation, and perception) from concerns about large language models, arguing that embodied robots are comparatively low-risk because physical reality is a built-in verifier — a robot's success at a physical task is directly and objectively checkable, unlike an LLM's factual claims — and that he's more worried about disembodied AI misinformation at scale than about robots "taking over." He discusses the company's co-authored open letter (with Unitree, Agility, ANYbotics and others) pledging not to weaponize their general-purpose robots, and his skepticism that current large language models possess anything like true consciousness or emotion (calling their operation "elaborate word completion"), while acknowledging Lex's counter-argument that statistically-derived outputs could still functionally simulate emotional connection. He closes on his own optimistic vision of a future where legged robots, like Sony's AIBO, serve as genuine companions addressing human loneliness — provided the underlying utility-first, safety-first engineering culture that built Atlas and Spot continues to scale.
Excerpts
On the core engineering philosophy inherited from Marc Raibert's lab — working with physics rather than against it:
“How do you build something so that the physics of the machine just kind of inherently wants to do what it wants to do... and figure out how to augment those natural physics with also intent — how do you then control that but not overpower it? It's that coordination that I think creates real potential — we could call it beauty.”
On why natural-looking humanoid walking took so long to achieve:
“We did build an early machine, we called it Petman prototype... it had a really nice looking gait... but even then it was hard to get the robot to walk where, when you're walking, it fully extended its leg... getting that all to work well took such a long time. In fact I probably didn't really see the nice natural walking that I expected out of our humanoids until maybe last year.”
On why he sees embodied robots as less risky than disembodied AI systems:
“If I'm building a robot to do a manual task of some sort, I can judge very easily: is it doing the task I asked it to, is it doing it correctly? There's sort of a built-in mechanism for judging... physical reality is a good verifier... whereas if you're asking a theoretical question in ChatGPT, it could be true or it may not be true, and it's hard to have that verifier.”