Humanoid Robot Funding Tops $11B as Foundation Models Advance

A month that showed the whole stack at once

Robots that combine vision, language and movement had a busy October 2026. In a single month, Tesla started building Optimus in Fremont, several labs released new robot foundation models, and venture money kept arriving at a record pace. Warehouse operators kept expanding automation as well, although most of it is less glamorous than a humanoid. Each development matters on its own. Taken together, they suggest that money and the AI "brains" of robots are improving faster than robots are actually being put to work. The near-term story belongs to the software layer and to robots that don't look like people.

The money: records everywhere, but nobody agrees on the total

Every tracker shows humanoid funding hitting records in 2026. They disagree on how big the record is. Dealroom counts $11 billion of venture capital into its humanoid category from January through October 6, more than twice the $4.8 billion it recorded for all of 2025.6 Tracxn reports $6.61 billion in equity funding across 42 rounds through September, up from $2.05 billion over the same period a year earlier.2 Humanoid500 puts disclosed 2026 funding at $45 billion. That figure is inflated by a $30 billion credit facility it assigns to Tesla's Optimus effort, so it is not comparable to the venture-only counts.4 The fair conclusion is that capital has at least doubled year over year, and the exact total depends on what you count.

The deals keep getting bigger and earlier. London-based Machenta raised a $40 million seed round co-led by Creandum and Union Square Ventures.37 Eka Robotics announced a $150 million Series A led by a16z and F-Prime but named no customers.39 Sequoia put $60 million into Mecka AI, which captures human motion data for training. Minerva Humanoids emerged from stealth with a $10 million pre-seed led by General Catalyst to build semi-autonomous humanoids for hazardous energy and bomb-disposal work.3 In China, CASBOT announced a Series A++ worth several hundred million yuan.10 Low-cost humanoid maker Noetix has reportedly filed for a Hong Kong IPO.32

Public markets are less enthusiastic. Unitree's shares fell nearly 40% in September from their first-day close, erasing about $35 billion from their peak value.38 Agility Robotics is going public through a SPAC merger with Churchill Capital Corp. It has Digit deployed at nine customer sites but reported just $1.8 million in revenue.38 When a company valued in the billions has a seven-figure revenue line, the market is pricing the future, not the present.

The brains: foundation models move past cameras

The biggest technical change is happening in robot foundation models. These are typically vision-language-action (VLA) systems that take camera images and an instruction and output motor commands.15 October brought a run of releases. Neura Robotics launched Neuron AI, which it calls a "Perception-Language-Action" model. It combines 3D vision, touch, force, audio, the robot's own state and language with memory, and it is trained across five of the company's robot designs.11 Galaxea published G0.5, which puts step-by-step reasoning and action generation into a single stream and replans at every frame.17 Google said its Gemini 4 Argon model taught a humanoid a throwing game from a video.17

Nvidia's open GR00T N1.7 model, licensed under Apache 2.0, added 20,000 hours of first-person human video to its pretraining to improve generalization. It also supports whole-body control on Unitree's G1.13 A lab called Nature Will announced Natura-1, a 50-billion-parameter model for dexterous hand control pretrained on more than a million hours of first-person human video. Coverage noted that it handles hands only, not whole bodies or walking.12 Machenta takes a different route with what it calls a vision-demo-action model. Instead of a language instruction, a single demonstration tells the robot what to do, and the model adapts without retraining.37

These efforts share a direction. Each is trying to get around the field's main bottleneck: synchronized records of what a robot saw, what it was told and what it did. That kind of data does not exist at internet scale.19 Human video, simulation, sensor suits and teleoperation are all being used to make up the shortfall.1113 Touch is a second theme. Eka trains what it calls a "vision-force-action" model,39 and HD Hyundai paired its in-house VLA with fingertip tactile sensing on a new dual-arm platform.19 Vision alone appears to have hit its limit for reliable manipulation.

The biggest AI labs now want both the brain and the body. Sam Altman has committed OpenAI to a humanoid demo in 2027, which puts the company in competition with Figure and 1X, firms it once backed.35 Boston Dynamics has partnered with Google DeepMind on AI for its new Atlas.38 My reading is that standalone model makers in the middle face a squeeze. Frontier labs are moving down into hardware, and hardware makers like Neura and HD Hyundai are building their own models.1219

The bodies: production starts, but volumes are small

Humanoid hardware is reaching factories in small numbers. Tesla said on October 21 that it had begun making what it called "production-quality" Optimus robots in Fremont. Musk warned of a "bumpy" ramp and still claimed 20,000 units a week by the end of 2027.39 Other reporting says Tesla built only a few hundred units against a stated target of 5,000.34 Figure had produced more than 350 Figure 03 robots by April and said its build time had dropped from one robot per day to one per hour.29

On global volume, two outlets cite roughly 19,100 humanoids shipped in the first half of 2026, up 272% year over year, mostly from Chinese makers such as Unitree, UBTech and AgiBot.3435 Toyota's entry, framed by the Financial Times as a race against Tesla and China, has so far produced no confirmed product details, price or production schedule.34

Industrial partners are making real commitments. Schaeffler opened a 7,500-square-meter humanoid facility in Taicang, China, where six humanoids build parts for other humanoids. It plans to deploy at least 1,000 AEON humanoids over seven years.29 Agile Robots told Reuters it will build 1,000 to 2,000 compact research humanoids next year that ship with Nvidia's GR00T software and Jetson Thor computing. Stanford and ETH Zurich are among potential buyers.32 At the low end, Faraday Future launched a $2,490 developer humanoid.33 A large share of the hardware being sold is going to researchers, schools and pilot sites rather than paying production lines.

The warehouse: still a job for non-humanoid robots

Warehouses show the gap between hype and use most clearly. North American companies ordered 17,995 robots worth $1.166 billion in the first half of 2026. That is up 6.6% by value but only 2% by units.30 Value growing faster than unit count suggests buyers are paying for more capable systems, not simply more robots. Amazon runs more than a million robots and says robotics help fill about 75% of customer orders. Its June 2026 Proteus accepts plain-language commands.30 An industry survey found that 52% of operations now use at least one type of warehouse robot, up from 48%.25

The systems doing most of this work are cube storage grids, mobile robots and picking arms. Brightpick robots complete 73% of picks without help at one Tennessee distributor, though people are still needed for loading and packing.25 Coverage also warns that automating one step often just moves the bottleneck to the next.25 One market estimate gives fixed automation such as storage systems and conveyors 53.2% of 2026 warehouse automation spending.28

Humanoids are entering this setting slowly. GXO will deploy Agility's new Digit 5 at sites in North Carolina and Tennessee starting in early 2027. Earlier Digit fleets logged more than 65,000 hours and moved over 100,000 totes.26 Agility chose a fast-charging system meant to keep Digit 5 working more than 20 hours a day.31 Even supportive coverage says today's humanoid deployments mostly involve moving totes, palletizing and delivering parts.24 Those are useful jobs, but they are not yet general-purpose work.

Where the reporting diverges, and what to make of it

The biggest disagreements are about numbers: total funding, Tesla's real output, and how "commercial" current humanoid deployments are. Optimistic reporting describes humanoids moving past pilots into limited commercial use.24 Skeptical newsletters call Tesla's milestone "a start, not yet a product." They point out that Unitree's promotional H2-X video gave no specifications and did not say whether the robot was autonomous or remote-controlled.39 Governance is lagging too. Biometric privacy lawsuits filed in Chicago against delivery-robot firms Coco and Serve claim that robot cameras and LiDAR capture people's faces, which both companies deny.39

My reading is that 2026 is the year robot intelligence became a serious, well-funded field of its own, and humanoid bodies are mainly a way to collect training data and win future markets. Money rewards the brain, and warehouse budgets still go to grids and wheeled robots. The test for 2027 is whether the new models can turn humanoid pilots into revenue at a pace that matches the valuations. Agility's $1.8 million in revenue is the baseline to beat.38