Three robot bets: where does the generalization ceiling come from?

Suppose you want a robot that can tidy a stranger's kitchen, one it has never seen. Robot people call that generalization, and three of the best-funded robot companies, Figure, Generalist and Physical Intelligence (Pi), all say it's the whole game. They disagree about where it comes from.

Figure builds a humanoid and pays people to film themselves doing chores. Generalist hands people a cheap two-finger gripper and records them using it. Pi trains one model on many different robots, then lets it practice on the job. Each is a bet on where the generalization ceiling is: the point where more data and money stop making the robot better in places it hasn't seen. Underneath sits one question: do you need a human-shaped robot to learn from human data? Figure says yes. If the answer is no, a simpler robot on wheels might be enough for a lot of homes.

Count your nines first

The idea of a ceiling comes from a talk by Dmitri Dolgov, Waymo's co-CEO, at Y Combinator's Startup School in July 2026: The Demo Is Only 1% Of The Work. Reliability comes in nines, 90%, 99%, 99.9%, and each one costs far more than the last:

"getting to that first 90% or 99%, that's the easy part. But then every next nine that you want to add, that takes about 10 times more effort. So you need to know up front exactly how many nines your product actually needs."

Every technology has a curve that starts steep and then flattens, and the classic mistake is "picking the tech that gives you the fastest early ramp". Cameras alone are "a very reasonable way to go" for a car that only has to drive like a person, but for superhuman driving "weak sensing just leads to a safety curve that flattens out way too early." In a dust storm his camera sees very little and the lidar shows a pedestrian by the road. More camera data can't fix that, because the pedestrian isn't in the pixels. His second example is learning: plain end-to-end imitation "will give you very rapid progress", but for a superhuman product "just doing kind of that basic vanilla end-to-end is not enough." I'll call that the imitation plateau.

So how many nines does a home robot need? Let's do a back-of-the-napkin calculation. Say it does 50 small tasks a day. At 99% per task it gets something wrong every two days. To fail less than once a month it needs about 99.93%. That's three nines at least, in homes nobody tested in.

Three bets on the ceiling

Figure thinks the ceiling is the body: people are the biggest source of physical know-how, and a robot gets the most from their data if it's shaped like them. Generalist thinks it's the data: one kind of data, real interaction with the action recorded in it, at enormous volume and fed to a big model. Pi thinks it's the mix: robot data from many robots, homes and tasks, with the robot's own practice on top. Figure reshapes the robot to fit people, Generalist reshapes people's tools to fit the robot, and Pi asks one model to absorb the difference.

All three are private companies.

Figure builds every layer

In Figure's All Hands 2026, posted on October 22, Adcock split the job into four chapters: robot hardware and the AI recipe, which he says "money can't buy", then scaling intelligence and scaling robots, which money can, and "you can't skip chapters." So Figure designs almost the whole robot itself, "including the actuators, motors, sensors, battery pack, and electronics" (Figure), and builds it in its own factory, which reached one Figure 03 per hour in April 2026 (Figure). It all serves the plan to learn from people. Adcock says they spent "an enormous amount of time" getting the robot's fingertips to track a person's "at millimeter level precision". You can only shape a hand around human data if you build the hand yourself.

Two ways to meet the data

Then Adcock gets to what he calls "the largest debate in robotics the next few years". His slide shows a fork. On the left, "design human collect for our robot": build specialized hardware for people to collect data with, then design the robot around it. He says "there are groups out there doing it now", that the road is fun and fast, "and it will have the lowest ceiling." On the right, "design our robot for human collect": record raw human behavior and build a body that can use it, "which means the body and the hands and the way the robot moves the world over time needs to match human movements." That road, he says, "has a much higher ceiling, maybe even unbounded ceiling."

He names no one, but the left road matches what Generalist does. Then he turns to robots that work at a table, again without names:

"And let me be clear, if you collect millions of hours on a table, you will generalize to a table and you'll pick up balls and do other stuff. [...] You have to generalize to that over there on the right. And then when you start generalizing to that, you're going to have to re-architect everything here on the left. I've been there. We've been on a table before, and we hit a wall."

He doesn't say table data is useless once the robot walks. He says leaving the table means rebuilding everything, and for Figure that meant a year of re-architecting Helix, its robot model, to send commands "to every joint". So the wall was partly in Figure's own model.

Is the humanoid the only way to win, then? Adcock says yes. Two claims are folded into that. One is that raw human data has a higher ceiling than data from robot-shaped tools. The other is that only a human-shaped robot can use human data, and that's the one that makes the humanoid necessary. So what has Figure tested?

Figure's test: 30 homes it never saw

A month before the all-hands, for Helix 2.5, Figure sent robots into 30 Bay Area homes "with zero data collected in any of them." The tasks were still taught with robot data from elsewhere, so "zero-shot" here means new homes and new objects. The model was "pretrained from random initialization entirely on Index", Figure's dataset of human video.

Then Figure ran a controlled comparison: the same task data, model and test, with and without Index pretraining, on three household tasks with 140 trials each. Folding towels went from 12 to 87 successes, picking up toys from 7 to 56, and making a bed from 16 to 94. Pooled, 8% became 56%.

That's the strongest public result from any of the three that human video makes a robot better in new places. But no robot without a human shape was in the comparison. It shows that human video helps Figure's humanoid, not that another body couldn't get the same lift.

Generalist: robot-shaped data from human hands

My prior on Generalist was that its edge is better pretraining from collecting much more data. That's partly right, but the data is a different kind than I assumed.

The volume is real: by April 2026 Generalist had "over half a million hours" (GEN-1). That isn't robot data, though. People hold "low-cost wearable devices", hand-held grippers with a camera, built to match the robot's own gripper. CTO Andrew Barry says it's "all about making sure the data set and the inference time matches" (podcast). The base models of GEN-0 and GEN-1 are "trained without any robot data", and the robot shows up only after pretraining, with about an hour of robot data per task.

The founders don't describe the edge as hours either. Pete Florence calls data "table stakes" and says "an hour of data is not a fungible unit" (podcast). Their pitch is a person's natural, fast motion, which a robot driven by remote control loses (Generalist), plus a model big enough to absorb it and tests of which data mixtures help most. So my corrected prior is one well-chosen kind of data at large volume, a model big enough to use it, and a fast way to learn which data matters.

Their results are about how little robot data a new task needs. With about an hour of robot data per task, GEN-1 reached 99% average success on three trained tasks, against 64% for GEN-0 and 19% for a model trained from scratch. An academic study with the same kind of device reached 85% to 92.5% on four tasks in eight unseen places with unseen objects, with no robot data at all (Lin et al.).

On the hand, Generalist disagrees with Figure: "Hands can change, physics does not" (Generalist). Florence still calls generalizing to new hands "not quite solved" (interview).

Physical Intelligence: one model, many bodies

Pi's bet is the least about one body. "The only answer to generalization that we know in machine learning is through diversity of data," co-founder Karol Hausman said in a Sequoia interview. In Pi's published models, most of that variety comes from robots.

The clearest case is π0.5, the 2025 model that cleaned homes it had never seen. About 400 hours came from its own two-armed robots on wheels in about 100 homes, and 97.6% of the first-phase training examples came from elsewhere: other robots, static arms in homes and the web. On Pi's chart, success with unseen objects was 94% with everything, 31% without the static arms in homes and 49% without the other lab robots.

Pi builds its own simple robots at a factory in Fremont (Pi): mostly two-armed, all with parallel-jaw grippers, on wheels or fixed in place. Sergey Levine says they avoid humanoids for data collection because "humanoids are expensive, complicated, and teleoperating humanoids is harder" (interview). Partners already run Pi's models on their own humanoids through its API. For Pi, the body is a variable.

The third part is practice. Pi's models are VLAs, vision-language-action models: robot policies built on a model of images and text. Pi puts Dolgov's imitation plateau in those terms: "it's relatively easy to get VLAs to succeed at a task some of the time, it's quite hard to make them succeed reliably" (Pi). Its method, Recap, trains on the robot's own attempts and on people's corrections, and "On some of the hardest tasks, RECAP more than doubles task throughput and roughly halves the task failure rate" (paper). In April, π0.7 folded that practice into one general model that matched the RL-trained specialists.

Human data on robots that aren't human-shaped

So do you need a human body to use human data? For the amounts tested so far, no. Human data gives measured gains on robots that look nothing like people. In Pi's human-to-robot paper, three hours of tracked human demonstrations raised a two-armed gripper robot on wheels from 32% to 71% on a spice-rack task, in a kitchen it had seen only in the human data.

Humanoids gain too: HAT +25 points, EgoHumanoid +51. The question is whether they gain more than other bodies, and the papers that compare bodies don't show it. ImMimic gave one arm four different hands and the same human videos. With the authors' method the grippers gained 26 points and the dexterous hands 25, and with plain co-training the hands gained more, 16 against 6. The authors find the gain tracks how close the human actions are to the robot's, "regardless of its mechanical structure". In EgoVerse, a Unitree G1 humanoid with hands gained in the same range as two gripper arms, on a shared 8-hour human dataset plus 2 hours recorded in each lab.

What hurts is misalignment. In the What Matters study, human video doubled success on one task, 26.2% to 51.2%, once the human camera's image scale matched the robot's, and added nothing with a badly mismatched camera. Scenes, objects, task order, routes and rough hand paths carry over to a gripper on wheels. Grasp style, finger contact and whole-body motion carry over poorly, and fine precision suffers because hand tracking is off by 1 to 2 cm (EgoZero). Whole-body motion is the one part only a humanoid can use.

One question stays open. These papers mix a few hours of human data into robot training, and Pi found that this helps only on top of varied robot pretraining. Figure pretrains on human video alone, at a far larger scale, and nobody has run that recipe on a robot without a human shape. Generalist comes closest: it pretrains a gripper robot with no robot data, but on hand-held grippers, not bare hands. Most of these tests also run 10 to 50 trials per condition, which can miss a 30-point gap between two bodies. The paper-by-paper evidence is in a companion article: Do you need a human body to use human data?

Legs, wheels and the flat with no stairs

Even if a robot on wheels can learn from human video, a humanoid can climb stairs and a wheeled robot can't. Isn't that the end of it?

Dolgov's frame gives both sides a point. A wheel at a stair is a ceiling of his kind, a physical limit that no data removes, and Figure says so: "wheeled robots need enough open floor for their base to fit and turn. Homes offer no such accommodation" (Helix 2.5). But counting nines starts from the product, the same way cameras alone are fine for a car that only has to drive like a person. For flats, warehouses and offices with elevators, the stair ceiling sits above the target, and what's left is cost and the hand.

That's the Roomba argument, and I think it holds. A Roomba doesn't climb stairs either, and it sold because it was cheap and good enough on one floor. On cost, Levine says each arm on Pi's robots costs about $3,000 (interview). Figure publishes no price, and Adcock describes Figure's next hand as having "more actuators than an entire robot".

So Figure's two claims get different answers. The body sets a hard ceiling on what a robot can reach, and legs raise it. That the body also limits how much a robot learns from human data is the claim nobody has shown.

Four ceilings, and who attacks which

I count four ceilings, and the three bets push on different ones.

• The body ceiling is what the robot can physically reach. It's the only hard physical limit, and only Figure attacks it.
• The data ceiling is what the data covers and what it contains. Human video covers the most and contains the least, with no actions and no forces. Generalist's grippers record the action, at the price of a robot with a pincer. Pi's robot data contains exact actions and covers less, so Pi mixes many sources.
• The demonstrator ceiling is Dolgov's imitation plateau: copying gets a robot close to the people it copies, and reliability takes more. Pi has the clearest answer, practice with corrections. Figure bets on large-scale pre-training over training task by task: "post-training is not going to win this game. It's pre-training that will win."
• The evaluation ceiling: you can't climb nines you can't measure. Showing a failure rate under 0.1% takes about 3,000 clean trials. The best published numbers sit at one nine or below in new places: Figure's 56% in new homes, π0.5's 94% with unseen objects, π0.7's 60% to 80% on unseen tasks (paper). Generalist's 99% is two nines, on trained tasks. Nobody has published a test big enough to show three.

Figure's question is the right one, and its answer goes further than its evidence. It has the best public evidence that human video helps a robot in new homes, and the body does cap what a robot can reach. But the claim that you need a humanoid to learn from human data is the part it hasn't tested. Video doesn't need the exact body. It needs a hand and a camera that line up with it. I also don't buy "the lowest ceiling" for Generalist's road: the hand-held results are among the best numbers here.

If I had to pick the highest ceiling, I'd pick whoever turns a robot's own experience into training data fastest, on a body chosen for the product. That's closest to Pi's bet. It's a bet, and the evidence is still thin: Recap's halved failure rate, π0.7 matching RL-trained specialists, and Dolgov on cars, "the actual experience of having your agent in the real world is invaluable and it's irreplaceable." In a house with stairs, the right body is probably a humanoid. In a flat, I think a simpler robot on wheels with a good hand wins, the way the Roomba did.

Here's the comparison in one place.

Sources and what isn't verified

Talks. Dolgov, The Demo Is Only 1% Of The Work (Startup School, July 2026), checked against two Whisper transcripts, YouTube's captions and YC's transcript; title from his Waymo bio. Adcock, Figure All Hands 2026 (posted 22 October 2026). That video has no captions, so its quotes come from two Whisper transcripts checked against each other. "[...]" marks a cut. The interview quotes (Barry, Florence, Hausman, Levine) are Whisper transcripts checked against YouTube captions or official transcripts.

Company basics. Figure: SEC Form D (Delaware, 2023), master plan (May 2022), careers (San Jose), Series C. Generalist: Form D (Delaware, 2024, San Francisco, officers), August 2026 Form D, funding post. Pi: March 2024 homepage, careers (San Francisco), Fremont job post, and CapitalG, which led the Series B and lists it at $600 million on its insights page.

Figure. BotQ, production ramp, Helix 2.5. The trial counts are from the post's Figure 1. Its text says 9%, and its counts give 35/420, which is 8%.

Generalist. GEN-0, GEN-1, GEN-1.5, physical commonsense, a thousand hands. Results are Generalist's own, without trial counts for the averages.

Pi. π0.5 paper and chart, Recap paper, π0.7 paper, human-to-robot paper (spice runs on its wheeled robots), API post. After π0, Pi gives only a rough data size, "tens of thousands of hours" (π*0.6). The $3,000 arm is Levine on the Dwarkesh podcast (2025).

Transfer papers. ImMimic: my averages of Table 1 over four tasks, 10 rollouts per cell. EgoVerse: read from Figure 9; three robots in three labs. What Matters: Table 5; its canvas dot is the low-data result, 12.0% to 41.7%. EgoZero: the 1 to 2 cm error. Other canvas dots: EMMA (82% against 52% for an hour of remote-controlled driving), EgoWAM (25% to 15% when the human grasp doesn't fit the gripper, 20 trials, within noise), HAT (59/170 to 101/170), EgoHumanoid (31% to 82%), LeLaN and ego-video navigation co-training. Tasks, labs and metrics differ.

Not verified. A Figure 03 price, Pi's founding date, and the valuations of Generalist and Pi.

Numbers. One failure a month at 50 tasks a day is 1 in 1,500, so 99.93%. With zero failures in $n$ trials, the 95% upper bound on the failure rate is about $3/n$. Detecting a gap between two bodies from a 50% baseline (5% test, 80% power) takes about 40 trials each for 30 points.