13-Hour Coffee and 2x Throughput: Physical Intelligence Crosses the Deployment Line

Karol Hausman opens with the thing that still startles him: you build a loosely brain-inspired learner, feed it data, and it “gets it” better than anything the field has had, for robots as well as vision, language, and sound. He is CEO of Physical Intelligence, the year-and-a-half-old company he describes in one sentence: robotic foundation models that, in principle, let any robot do any task. Tobi Springenberg, his colleague from Google’s robot-learning era, is there to talk about the latest release. Sequoia’s Sonya Huang and Alfred Lin host.

Over eighteen months they have shown the building blocks of scale: one model driving many robot form factors, some generalization into rooms the robot has never seen, and now a performance jump they think is finally deployable. The model is π*0.6. On the company’s site, the same weights make espresso for thirteen hours in the office and fold laundry for four, recovering mid-run instead of being reset. They also fold chocolate boxes for a company that actually ships product, at more than double the old throughput. Hausman had assumed commercially valuable work was about five years away. They did it, he says, roughly two months before this taping. “I think we just crossed that threshold.”

Deployment matters twice. It is the product. It is also the data. There is no free internet of robot actions, so PI has to create the datasets itself. Hausman’s bet is that robots doing economically useful jobs will be the cheapest collection method of all — cost of data, in the limit, negative — and that this stream will eventually dwarf the internet. The company does not build bodies. It builds the brain, on the theory that intelligence was always the missing piece.

Huang presses the obvious competitor set: household helpers, cooking robots, humanoids, fully vertical products. Why a foundation model? Hausman says the history of robotics is unambiguous to people who work in it. The field has always been bottlenecked on intelligence. More than a decade ago there were robots that, if teleoperated, could clean an entire house. The caveat is the whole point. Put a human mind behind the hardware and the hardware is already capable of a lot, at home and on the factory floor. Companies still organized around a single task, he argues, are not attacking the constraint.

Huang pressure-tests the “why now” with Tesla Optimus’s new hand, which she calls a piece of art, and with tasks like peeling and dicing an onion that may only now be physically possible. Hausman grants the progress, especially in dexterous hands. It still does not resolve the bottleneck. You could have chopped vegetables with simple grippers years ago if anyone had the intelligence to run them. More complex hardware raises the ceiling of what a robot *could* do. It also raises the intelligence required, which is the floor the industry has not reached. Huang’s summary, which he accepts: hardware lifts the ceiling; the capability floor is not there yet.

If they crack the intelligence layer, he says, vertical products will grow underneath it — homes, industry, “basically anywhere.” That is why PI aimed at the bottleneck instead of shipping a cute helper of its own. Until a model can operate even a simple robot at human-teleoperator level, prettier hands are not the product.

Lin keeps drilling: if intelligence is the limit, what limits intelligence — data, cost, generalization? Hausman splits the work into three factors. Capability means that if you can collect data for a task or a robot, the model should be able to replicate it. They got there fairly quickly with π0 about a year earlier: many robots, many tasks, one model, as long as the demonstrations exist. Generalization is still open. Drop the robot in a new home and it has to know where things are, how the counters look, how the light falls. The only answer machine learning currently has, he says, is diversity of data. π0.5, released in April of that year, could operate in a home it had never visited — not perfectly, but with enough common sense for jobs like cleaning a kitchen.

The third gate is performance: fast enough, reliable enough, long enough that you would actually leave the robot running. That is what π*0.6 is for, and why they have started deploying themselves. How wide is the aperture? Hausman will not pretend to know. Foundation models, like language models, are cooked in-house and only then tested; you cannot fully predict the artifact. So they open-sourced the weights and let other people try. The community put them on driving, surgical robots, and agriculture — farther than he had expected. The aperture, he thinks, will keep growing with data.

Springenberg adds the cold water. The starting point is wider than they thought. Getting each of those starting points to a level where someone would run a business on it every day is still a lot of hill-climbing. The long tail of things that can go wrong is not tamed. π*0.6’s contribution is learning from experience after deployment so the model can get better on the job. That, he says, is progress, not a finished grasp on every failure mode.

The star in the name nods at Q-star; Springenberg says they are really aiming at policy star, the optimal policy. Until this release, PI’s foundation training was imitation: humans teleoperate, the model copies. π*0.6 adds RL from experience the robot collects by running. You start from the demonstration policy, deploy it, let humans label success and failure, and let them intervene with corrections. That stream comes back so the model can reinforce some behavior and suppress the rest. The missing piece, Hausman had said, was how to escape the plateau you hit when you simply collect more of the same demos. This is that piece.

They did the RL in the real world, not in sim. Sim is a research tool; the π*0.6 paper’s learning happened on physical systems, because the long tail does not live in a tidy simulator. One deployment task is folding cardboard chocolate boxes for real outbound orders. A new shipment arrived imperfectly perforated, sheets stuck together. The robot grabbed one, put it on the table, and suddenly had two boxes. A simulator written to hand over single sheets would never have shown that. Springenberg’s point is not the anecdote. It is that you cannot sim-then-deploy your way past the failures that only show up when the world is sloppy.

Hausman offers a structural reason sim-to-real worked for walking and not for hands. Locomotion is mostly modeling your own body; do that once, for one robot, and transfer can work. Manipulation is how the world answers you. Moving a hand from A to B is easy. Predicting what every object, in every task, will do is not. That is why those methods have not scaled for picking and folding. On the three tasks they cared about — box building, an industrial espresso machine, and laundry — they more than doubled throughput, and the robots recovered instead of dying on the first mistake. Watching it live, Springenberg says, changes whether you believe deployment is a toy demo or the real thing. Robotics has never lacked cool videos. You can take as many takes as you want. A robot that fails every other trial is not deployed.

Two details stuck. On espresso tamping, the robot at first pressed so hard it almost lifted itself off the table, because the human demos only cared that the puck was flat. After 30 to 50 correction episodes — tiny against millions of pre-training runs — it became gentle. Springenberg was shocked that so little data could turn the behavior. The other is the value function, trained on the same backbone, which predicts from the current step whether the episode is heading for success. With more tasks in the mix it got good enough to sour 30 to 40 steps before a bad portafilter angle was obvious to a watching human. Huang hears a rebuttal to Andrej Karpathy’s “slurping bits through a straw”: the robot is not waiting for a terminal reward. Hausman replies that RL is a problem definition, not a synonym for policy gradients, and that value functions and temporal-difference learning exist because robotics has no perfect language simulator to brute-force.

Hausman compresses the field into four generations. First, handwritten rules: throw enough engineers at the world and you can code every case. The world was too complex. Second, a modular pipeline — perception, planning, control — that grew three academic communities with their own conferences. Learning each module separately beat rules, then broke at the predefined interfaces: perception emits a pose, planning emits a trajectory, control executes, and information dies at the seams. “When I try to pick up this glass, I don’t think about it in terms of perception and then planning and then control. I just pick up the glass.” Third, end-to-end: pixels in, actions out, let the net invent its own internals. The direction was right and the data hunger was fatal; first-person action data cannot teach every piece of common sense. Fourth is the VLA era they are in now: an internet-pretrained vision-language backbone, plus an action expert that actually drives the robot.

Today it is one model, up to about seven billion parameters, images and text in, text and actions out. Most of the training mix is now their own robot data; internet data is a minority. Springenberg would not be surprised if, in five or six years, the VLM backbone has been swapped out. He thinks the data-and-training *practice* will last. The models already plan roughly the next 50 steps, a second or two of motion, and they decompose “clean the kitchen” into language subtasks — drive to the counter, pick up the glass, put it in the sink. He does not think physical reasoning will look like math RL. Learning tennis, he says, is not inner speech about the racket; it is motion and the trajectories of objects.

Hausman goes further. Right now robotics benefits from VLMs. He expects the flow to reverse. Many LLM shortcomings are baked in by text tasks like math and code. Robotics will force reasoning in a mixed abstract space — some text, some images, some trajectories — grounded in a physical world that does not give you infinite rollouts. Those methods, he thinks, will flow back. On continual learning, Huang cites Rich Sutton’s era-of-experience thesis. Hausman says the models are more “alive” than a frozen pretrain-plus-post-train artifact: they keep trying things after deployment. They are not classical continual learning. Springenberg was still shocked at how fast small corrections land. Learning to tamp, he says, does not make you better at folding boxes. The mechanism is reusable: thirty corrections here, thirty there, and the pile lifts overall skill. True cross-task generalization still comes from pre-training, and from the flywheel of deployed data coming home.

Some people want to stuff Newtonian physics into the weights so the model does not have to learn it. Hausman has watched that fail. Hard-coding known rules limits what the model can learn next. His analogy is animals. A deer is about as smart at birth as it will ever be. Humans and crows have long, clumsy childhoods and earn intelligence from experience. There is no preinstalled version. Machine learning, he says, is paying the same tuition.

Huang asks what they sell: a whole robot, or a model customers have to integrate? Hausman’s answer is the most honest three words in the hour: they do not know yet. The technology is only now at the deployment threshold. Robotics startup history, as he tells it, is a trap. You begin with a grand general-purpose vision. The moment you pick one application, you cut corners into special-purpose hacks, and soon you are a warehouse pick-and-place company. PI wants to dodge that loop by making the technology as general, and the aperture as wide, as possible, and only then figuring out whether they are a model provider, a vertical product, or something they cannot see yet. Picking now would comfort the investors. He thinks it is too early.

Lin’s spicy follow-up is the shadow of self-driving. Stanford won the DARPA challenge in 2007. Sebastian Thrun was on a TED stage around 2009. In 2025 Waymo still cannot casually take the freeway. Why won’t general robots wait another fifteen years? Hausman splits the comparison. Easier: many tasks are fine at 95% reliability. A home robot that botches every hundredth folded shirt is acceptable; a car that fails catastrophically every hundredth trip is not. They also have foundation models with common sense, a different era than 2009. Harder: this is not one application. It is driving and manipulation and flying at once.

What gives him hope is an empirical surprise. The same model, fed mixed data, seems to benefit across driving, making coffee, flying a drone, and running a surgical robot — jobs that look unrelated enough that they “should” have been different systems. There is something about physical intelligence, he says, that they do not fully understand, and it might make the problem *easier* than the Waymo timeline implies. He does not want to draw the wrong lesson from cars.

Asked what has impressed them *outside* their own results, Hausman picks video models: the improvement curve was steeper than he expected, and the outputs are now hard to tell from reality. Springenberg is still in awe that next-token prediction produced systems that win math olympiads and help with scientific discovery. At the beginning of the year it felt as if LLM pre-training might be petering out. By year’s end it had taken a second breath. Hausman comes back to his opening: stop and notice that this loosely brain-inspired general learner, fed data, just works, on robots and on everything else. A robot that knows what to do in a house it has never entered, or that makes coffee for thirteen hours, is that same object. They do not fully understand it. It is starting to get it.

Lin notes that science taught everyone to smash big problems into small ones, and that someone later noticed this is a bad way to train machines. Springenberg says the whole ML field made that mistake: deep specialists, then a dream of multitask learning, and then the shock that a general pre-training objective made the pieces fall out. He does not see the accordion swinging back. Hausman agrees. “Best of both worlds” — bake in the rules we already know — has not worked in their hands. Let the algorithm learn from data. Huang calls the mission beautiful. The practical residue of the hour is less mystical: watch the open weights, score any embodied-AI company on capability then generalization then un-fakeable long-horizon performance, and treat every deployment as a data engine whose cost of collection can go negative.