The Humanoid Robot Landscape in 2026, Mapped by Data

Count the humanoid startups that announced a foundation model in the past eighteen months and you run out of fingers before you run out of names. Figure, 1X, Physical Intelligence, Skild AI, Apptronik, Agility, plus the research groups at NVIDIA, Google DeepMind, and Toyota. The hardware is quietly converging: two arms, two legs, a stereo camera or three, hands that are dexterous or trying to be, a battery good for something like a shift. Watch enough launch reels and the machines blur together.

What does not converge is where the training data comes from. That is the axis worth mapping in 2026, because a humanoid is only as capable as the demonstrations it has seen, and every serious team has placed a different bet on how to get them. Teleoperate a fleet? Synthesize motion in simulation? Mine human video? Pool trajectories across labs? Each answer carries its own cost structure, its own failure mode, and its own ceiling.

So this is a field guide organized by that one question. Not "who has the best robot," which nobody can honestly answer yet, but "who is feeding their robot what," which tells you a great deal more about who is still standing in 2028.

Four ways to feed a humanoid

Strip away the branding and almost every program in the field draws from one of four data sources, usually blending two of them. You can teleoperate physical robots and record clean action labels. You can generate motion and pixels in simulation. You can borrow first-person human video, which is abundant but carries no robot actions. Or you can pool everyone's robot logs into a shared corpus and train across embodiments. The table below is a rough map, not a scorecard: most teams hedge, and the interesting detail is which source they lean on hardest.

The on-robot fleets: perfect labels, brutal economics

The purest strategy is also the oldest: put a human in the loop, let them puppet the robot through a task, and record every joint angle and camera frame. The action labels are exact because the robot really did the motion. Figure runs this playbook with its own operators, and Toyota Research Institute has published Large Behavior Models trained on teleoperated diffusion-policy demonstrations, a single network absorbing hundreds of skills. Physical Intelligence went further with π0, a flow-matching policy trained across a fleet of different robot bodies, on the order of ten thousand hours of manipulation, so that one model transfers to arms it was not born on.

The weakness is arithmetic. Teleoperation runs in real time at roughly one operator per robot. An hour of data costs an hour of skilled human attention, and the rare, awkward, recovery-from-failure moments that policies most need are exactly the ones operators produce least often.

Teleoperation buys perfect labels at the worst possible unit economics: one human, one robot, one hour, for one hour of data.

Simulation and the world-model bet

If real data is slow, make fake data fast. Simulation renders millions of trials in parallel, with free and exact labels for every contact and joint. NVIDIA's GEAR lab built Isaac GR00T around this idea, releasing open weights for GR00T N1 and N1.5 trained on a heavy mix of synthetic trajectories and human video, so smaller teams can fine-tune rather than start cold. Skild AI leans harder still on simulation, pitching an embodiment-agnostic robot brain meant to drop onto many machines.

The catch is the reality gap: a policy that is flawless in sim can still fumble a real cable or a translucent cup the renderer never quite modeled. World models are the current answer. 1X trains a learned neural simulator of its home environment, using it both to generate experience and to score policies before they touch a physical robot. The bet is that a good enough predictor of the next frame is worth more than another warehouse of teleoperators.

Borrowing human video

Every person doing dishes is an unlabeled demonstration of dexterous manipulation, and there are billions of them. That is the appeal of first-person human video. Academic sets like Ego4D and its paired follow-up Ego-Exo4D captured thousands of hours of egocentric activity, and this footage increasingly seeds humanoid pipelines: GR00T's training mix draws on human video, and Gemini Robotics inherits the visual and semantic priors of a web-scale model before it ever sees a gripper.

The problem is the embodiment gap. Human video shows what happened but not the joint torques, contact forces, or gripper commands that produced it, so the actions must be inferred, retargeted, or paired with a small slice of robot data to become trainable. Abundant supply, missing labels: the mirror image of the teleoperation problem.

The shared commons

Open X-Embodiment pooled more than a million real robot trajectories from twenty-two embodiments across twenty-plus labs, and showed that a policy trained on the mix often beats one trained on any single robot's data. DROID added roughly seventy-six thousand demonstration trajectories gathered in the wild across dozens of buildings. Hugging Face's LeRobot has become the open plumbing that ties these together: standard dataset formats, pretrained policies, and a hub where community datasets accrete.

Cross-embodiment training is now a default assumption rather than a research curiosity. It also quietly reframes the competition. If the architectures are converging and the base datasets are shared, the edge moves to whoever can add data the commons does not already contain.

What to watch as 2026 turns

The honest read is that no single data strategy has won, and the leaders are the ones running two or three at once: teleoperation for clean labels, simulation for volume, human video for diversity, shared corpora for coverage. The question for the next eighteen months is less "whose robot walks best" and more "whose data pipeline compounds." Watch where the marginal demonstration comes from, what it costs, and who is allowed to use it. That is the map that will still be legible in 2028.