Robot Foundation Models in 2026: 500,000 Hours of Human Video, 50-Step Action Chunks, and a Winning Score of 0.26

I run compute for agents that act on software. The agents that act on the physical world are the same architecture with a harder output layer, and the labs building them raised something like $5 billion in the last twelve months, so I spent this week reading their papers, their filings and their model cards to understand what they have actually built. This is that read, as of 8 September 2026. It is more sceptical than the funding suggests and more impressed than the failure videos suggest.

What a VLA is

A vision-language-action model takes camera images, the robot's joint state and a text instruction, and outputs the next few motor commands. The architecture that won is a vision-language model with an "action expert" bolted on. Physical Intelligence's π0 paper, the reference design, uses a 3-billion-parameter PaliGemma as the backbone and a separately weighted 300-million-parameter action expert; the expert generates a chunk of 50 actions at once, one second of motion at 50 Hz, by conditional flow matching, ten Euler integration steps from noise to a trajectory. The backbone's key-value cache is computed once per observation and only the action tokens are recomputed per integration step, which is what makes it fast enough.

Two older ideas do most of the work. Action chunking, from the 2023 ACT paper, predicts a sequence of actions per inference rather than one, which fights the compounding error of behaviour cloning; ACT got 80 to 90% on six real tasks from ten minutes of demonstrations. Diffusion Policy, also 2023, generates that sequence with a denoising process instead of a regression, which handles multimodal demonstrations (there are many correct ways to grasp a cup) and gave a 47% average improvement across twelve tasks. Flow matching is the faster cousin of diffusion. Everything since is a bigger backbone, more data, and a better way to handle latency.

Latency is the constraint software agents do not have. A one-second chunk means the model has one second to produce the next chunk, and in Physical Intelligence's real-time chunking work the measured budget for remote inference was 139 ms: 97 in the model, 21 on the network, 11 resizing images, 10 elsewhere. Their trick treats the hand-off between chunks as an inpainting problem, freezing the actions that will execute during inference and generating only the rest, and it kept throughput flat with 200 ms of injected delay where naive inference degraded. Figure's Helix 02 stack goes further and splits the problem into three rates: a semantic reasoning tier, a 200 Hz pixels-to-whole-body policy, and a 1 kHz learned balance controller that they say replaced 109,504 lines of hand-written C++. That is roughly the split between planning, control and reflex in a nervous system, and it is where these systems converged for the same reason.

The data ladder

The reason the architecture matters less than it used to is that the labs have discovered the same scaling behaviour as language, and the argument is now about data. The ladder, roughly:

The GEN-1 result is the one I keep turning over: pretrain on half a million hours of humans doing things with instrumented gloves, then about one hour of robot data per task, and they claim 99% versus 64% success and roughly three times the speed of the robot-data baseline, 86 consecutive T-shirt folds, 1,800 block-packing iterations. NVIDIA's N1.7 makes the same bet with a shared relative-end-effector action space for humans and robots. If it holds, the data bottleneck moves from robots, which are expensive and few, to people wearing sensors, which are cheap and many, and the economics of the field change. It is also a claim from a company that raised money on it, with no third-party replication I could find, so hold it loosely.

I wanted to see the data rather than describe it, so I installed LeRobot 0.6.1, Hugging Face's open robot-learning library, and pulled two datasets' metadata from the hub:

from lerobot.datasets.lerobot_dataset import LeRobotDatasetMetadata
for repo in ["lerobot/svla_so101_pickplace", "lerobot/droid_1.0.1"]:
    m = LeRobotDatasetMetadata(repo)
    print(repo, m.total_episodes, m.total_frames, m.fps, m.robot_type)

lerobot/svla_so101_pickplace | episodes: 50 | frames: 11939 | fps: 30 | robot: so100_follower
  action: float32 (6,)   state: float32 (6,)   cameras: up, side (480×640 video)
  task: 'pink lego brick into the transparent box'
lerobot/droid_1.0.1       | episodes: 95658 | frames: 27630375 | fps: 15 | robot: Franka
  action: float32 (8,)   state: float32 (8,)   cameras: wrist_left, exterior_1_left, exterior_2_left
  tasks: 'Put the blue block in the green bowl', 'Move the bottom right tip of the duvet to the left', …

That is the whole modality in two lines of output. A robot is a 6- or 8-dimensional float vector at 15 or 30 Hz plus two or three video streams, and a "task" is a sentence. The first dataset is fifty episodes of a $200 3D-printed arm, the SO-101 with six Feetech servos, picking up a Lego brick; it is the dataset the SmolVLA paper was fine-tuned on. SmolVLA is 450 million parameters, trained on 481 community datasets totalling 10.6 million frames for about 30,000 GPU-hours, and it matches π0 on the LIBERO simulation suite, 87.3% against 86.0%. The DROID metadata fetch is 27 million frames from a $30,000 Franka arm across 13 universities. Both load with the same class. LeRobot's 0.6 releases in July and August added GR00T N1.7, π0.5, X-VLA and a RoboCasa365 benchmark, and its policy directory now has twenty entries; I count it as the most important open artefact in the field, because it is the only place a person with a laptop can run the same code as the labs.

Who shipped

Money first, because it explains the noise. Figure raised over $1 billion at $39 billion last September. Skild AI raised $1.4 billion at over $14 billion in January on 2025 revenue of about $30 million. Physical Intelligence raised $600 million at $5.6 billion in November and was reported in talks at over $11 billion in March. Apptronik's Series A passed $935 million in February. Agility is going public by SPAC at $2.5 billion on reported revenue of about $1.8 million. Generalist, Dyna and a dozen others raised nine figures. These are language-model valuations for companies whose products are, in most cases, pilots.

Now units. Unitree's IPO prospectus on the Shanghai STAR market, which listed on 19 August after raising about $905 million, says the company shipped more than 5,500 humanoids in 2025 and that humanoids went from 1.9% of revenue in 2023 to 51.8%. Its store today lists the R1 from $4,900, the G1 at $13,500 and the H2 at $29,900. AgiBot, filing for a Hong Kong listing, reports 15,000 cumulative robots by June. The two analyst counts of first-half 2026 global humanoid shipments, Counterpoint at 22,000 and Smart Analytics Global at 19,100, agree that AgiBot and Unitree are three-quarters of the market and that the full year will be 50,000 to 60,000 units. Counterpoint's breakdown is the part worth reading: over 60% of those shipments went to entertainment and to data-collection or research, 13% to industrial use and 5% to logistics. Most humanoids sold so far are being bought to generate training data for humanoids.

The American story is deployments rather than units. Figure 02 supported over 30,000 BMW X3 bodies at Spartanburg in 2025 doing sheet-metal insertion, and Figure 03 started a parts-sorting task there in June; neither company states the fleet size, and the "40 paid units" figure circulating is in neither press release. Agility's Digit has 65,000 operating hours and 100,000 totes moved across nine customer sites. Boston Dynamics and DeepMind announced at CES that they are building VLA models for the production Atlas, every 2026 unit of which goes to Hyundai and DeepMind, with Hyundai targeting 30,000 units a year of manufacturing capacity by 2028 and around 25,000 deployed in its own plants. Google DeepMind's Gemini Robotics 2, released on 30 July, is the first of the frontier-lab models to do whole-body control, walking, crouching and manipulating on Apptronik's Apollo 2, and adapts to a new body "in a few hours"; the reasoning model is publicly available and the VLA is trusted-tester only. NVIDIA's GR00T N1.7 is the open one: 3 billion parameters, Apache-2.0 code, weights on Hugging Face, and a Jetson Thor to run it on, at 2,070 sparse FP4 teraflops and 128 GB for the top part or $1,999 for the T4000 at volume.

And then Tesla. The January 2025 target was 10,000 Optimus units in 2025; in January 2026 Musk conceded that none were doing useful work. The Model S and X line at Fremont was torn down in May to make room, the first-quarter call said production would start "late July or August" and be "quite slow" with 10,000 unique parts, and the second-quarter letter in July said the lines were still being installed. The 10-Q for the quarter does not mention Optimus. I could not find a primary source that production has started.

The teleoperation problem

Most of what you have seen a humanoid do on video was done by a person. The November 2024 Optimus hand demo was confirmed on stage as teleoperated; the We, Robot event was "human-assisted"; the Miami demo in December fell over and the operator's headset-removal gesture was visible. 1X, whose NEO opened preorders at $20,000 or $499 a month for deliveries "in 2026," is unusually honest about it: the product ships with scheduled sessions where a "1X Expert" in a VR headset does the task remotely, and the CEO's line was that "if we don't have your data, we can't make the product better." Their Hayward factory opened in May with capacity for 10,000 units a year. I could not confirm a single customer delivery as of today.

I do not think teleoperation is a scandal. It is the data-collection strategy, and it is the same one the language labs used when they paid contractors to write demonstrations. The problem is only that the demos are presented as autonomy, and the benchmarks say otherwise. The strongest number I found is from Stanford's BEHAVIOR challenge: the 2025 winner, a policy based on π0.5, the best model in the field at the time, scored about 0.26 on fifty household tasks. The 2026 challenge, running now with 100 tasks and 1,950 hours of demonstrations, uses π0.5 and GR00T N1.7 as its baselines. DeepMind's own model card for the on-device version of Gemini Robotics 2 reports 53% on the SO-101 arm, up from 7% for the previous version, and says multi-finger dexterity "remains challenging." A factory task with one part, one fixture and a fixed camera is not a kitchen, and the benchmark gap between them is the gap between the BMW deployment and the home-robot promise.

There have also been injuries. The OECD's incident monitor records an elderly spectator hurt by a humanoid in Chengdu in February, an XPeng robot falling on stage in Shenzhen the same week, and a Unitree G1 that kicked a child in Xinjiang in June; the control mode was unclear in each case. A 30 kg machine that can lift 70 kg and is sometimes driven by a person over the internet is a new category of thing to have in a house, and the safety engineering for it is behind the model engineering.

What I take from it

The architecture is settled and open: a 3B VLM, a flow-matching action expert, 50-step chunks, and a real-time controller underneath, all of it reproducible in LeRobot on a $200 arm. The data is the moat and the labs are racing to collect it by every means available, including selling robots at cost to people who will generate it and paying humans to wear gloves. The deployments that work are single-task, single-fixture and industrial, and those are real and growing. The deployments that are promised, general household help, sit at a quarter score on the honest benchmark. And the valuations price the second thing while the revenue comes from the first.

For someone in my line of work the interesting question is where the compute goes. A humanoid fleet of 30,000 at Hyundai is 30,000 inference endpoints with a 139 ms budget and a body that falls over when the endpoint stalls. That is an edge problem, not a cloud one, which is why NVIDIA put a 128 GB Jetson in the robot, and why the on-device model card is the document I would read most carefully. The robots that ship will be running the same serving architecture I described last week, on a battery, with a fall as the failure mode instead of a timeout.

Related: LLM Inference in 2026, Neuromorphic Computing in 2026 and It's 2 AM. Do You Know What Your AI Agent Is Doing?.