One brain, many bodies

Our flagship piece made one argument above all: the binding constraint on physical-world AI is not architecture and not compute, but data that doesn't exist yet and has to be captured by hand, one expensive episode at a time. If you accept that, a hard question follows immediately. When every demonstration costs a human wage and an hour of someone's day, what is the single most wasteful thing you can do with it?

Train one robot.

That sounds glib, but it's the whole game. The default in robot learning has been to collect data on a specific arm, for a specific task, and train a policy for that arm. Every episode improves exactly one machine. Spend ten thousand operator-hours and you have taught one body to do a handful of things. The data wall isn't just that capture is slow — it's that the field keeps spending its scarce examples on an audience of one. Cross-embodiment learning is the bet that you don't have to.

Pool the data, not the hardware

The idea is simple to state. Instead of training one policy per robot, you train a single policy across many robot bodies at once — different arms, grippers, mobile bases, even legs — on their pooled demonstrations. Data captured on body A helps body B. The bet is that motor skill, like language, has a large shared core that a single model can learn once and reuse, plus a thin body-specific shell that's cheap to adapt.

Why would that be true? Because manipulation shares deep structure across embodiments. The physics of contact — friction, force, the moment an object starts to slip — is the same whether the gripper is two fingers or five. The geometry of grasping a mug, the sequence of reach-align-close-lift, the way a task goes wrong: these don't belong to any one robot. They're properties of the world the robot acts in. A foundation model over actions can absorb that shared part from everyone's data, then specialize the last mile to whatever body it's driving. It's the same wager that made language models general — learn the structure once, at scale, from a pooled corpus — pointed at motor control instead of text.

You can write the bet down. Let \(\theta\) be the parameters of the policy. Cross-embodiment learning assumes they split into a large shared block \(\theta_{\text{shared}}\) — contact, geometry, task structure, learned from everyone's data — and a small per-body block \(\theta_{\text{body}}\):

If \(\theta_{\text{shared}}\) dominates — if most of the skill is the part that transfers — then every episode anyone collects, on any robot, pays down the cost of the part that matters most. That's the entire promise: capture once, useful many times.

The evidence it's not just a nice idea

This would be a thought experiment if nobody had pooled data at scale. They have. Open X-Embodiment is the proof of concept: sixty-plus labs pooling their robot demonstrations into one corpus spanning twenty-two distinct robot types — the largest cross-embodiment dataset the field has.¹ It exists precisely because the single-robot approach hit a wall, and the only way past it was to share.

The headline result is the one that matters for the thesis. Models in the RT-X line, trained on the pooled cross-embodiment data, outperform the same models trained on any single robot's data alone — including on the very robots whose data was in the pool.¹ Adding other robots' experience didn't dilute performance; it raised it. That's the empirical signature of a real shared structure: the model learned something from the humanoid's data that made the arm better, because underneath, they were doing the same physics.

And the field has noticed. Robot-learning and robot-foundation-model papers grew more than 60% from 2022 to 2024, outpacing every other applied-AI subfield.² A lot of that energy is pointed straight at the pooling question — how to fold more bodies into one model, and how much transfers when you do.

The honest limit — the embodiment gap is real

Here's where I'd push back on the optimists, including the optimist in me. Transfer is real, but it is partial, and the reason is the embodiment gap: a 7-DOF arm and a five-fingered humanoid hand do not share an action space. Their joints, their reach, their contact geometry, the very dimensionality of what they can command — none of it lines up. A trajectory that's optimal for one is meaningless, sometimes physically impossible, for the other. The shared structure I keep invoking lives under the hardware; the hardware itself still differs, and that difference doesn't transfer for free.

So pooling is a multiplier on captured data, not a substitute for capturing it. Fold a new, weird body into the corpus and you still need real episodes on that body to pin down its \(\theta_{\text{body}}\) — the pool just means you need far fewer of them, because \(\theta_{\text{shared}}\) came for nearly nothing. The honest way to read the RT-X result is not "data is free now." It's "the same episode is worth more, because it's working for the whole fleet instead of one arm." That bends the cost curve our second pillar was about — it does not repeal it. You still have to show up and capture.

Why this is the most hopeful pillar — and the most concentrating

If you're worried about the data wall, cross-embodiment is the genuinely encouraging part of the story, because it means the wall is scalable-down: the field can share its way to more leverage per episode. Each new body added to the pool makes every other body a little better. That's the closest thing physical AI has to the compounding the internet handed language models for free — except here, the corpus has to be deliberately assembled, demonstration by demonstration.

Which is exactly why it cuts the other way too. A pooled corpus is a flywheel, and flywheels concentrate. Whoever controls the biggest, best-pooled dataset trains the strongest shared model; the strongest model wins more deployments; more deployments feed more bodies and more episodes back into the pool. Open X-Embodiment is open, which is a real and deliberate counterweight. But the same logic that makes pooling powerful makes a private, well-pooled corpus a moat — and the incentive to keep yours closed grows with exactly how well pooling works. The hopeful route and the concentration risk are the same mechanism, read twice.

What would make me wrong

Two ways. The first: the embodiment gap is wider than the optimists think, and transfer plateaus. Maybe \(\theta_{\text{shared}}\) is smaller than we hope — maybe so much of real-world skill is body-specific that pooling buys a one-time bump and then flattens, and a model spread across twenty-two robots ends up a generalist that's mediocre on all of them rather than strong on each. The RT-X results are early and run on a particular slice of tasks; "transfer helps" at this scale is not yet "transfer keeps helping" at ten times the bodies. If the curve bends back down, pooling is a useful trick, not a path out of the data wall.

The second cuts the opposite way: synthetic data makes pooling moot. If world models get good enough to generate action data on demand, you don't need to pool scarce real episodes across bodies — you mint as many as you want for whatever body you like, gap and all. Then the moat isn't who pooled the most capture; it's who has the best generator. I don't think we're there — synthetic data still has to be grounded in and validated against the real thing, and the embodiment gap shows up in simulation too. But it's the bet that would retire this one.

For now the shape holds. Every episode is precious, so don't spend it on an audience of one. Pool it, let the shared structure do its work, and accept that the body-specific last mile still has to be paid for in the real world. Cross-embodiment is how the field gets more out of each expensive recording — a multiplier, honestly bounded, on the one input that was never going to be free.