What is Hugging Face doing in robotics? LeRobot, Reachy Mini and the Pollen deal explained
Hugging Face's robotics business is an open-source software stack, LeRobot, plus cheap hardware from Pollen Robotics, which it bought in April 2025: the aim is to become the place where robot policies and datasets are shared, not to out-build Figure or Physical Intelligence.
LeRobot had 28,209 GitHub stars and 5,893 forks on Oct 16, 2026, and its July release added world-model policies, NVIDIA's GR00T N1.7 and a deployment tool that turns robot failures into training data.
Hugging Face sells a $399 desktop robot and a $70,000 humanoid, but neither is the point. The company's robotics business is LeRobot, an open-source library for recording robot data, training policies and running them on real machines, and the hardware exists to put more people on that software. On October 16, 2026 the LeRobot repository on GitHub had 28,209 stars and 5,893 forks, according to GitHub's public API, which makes it one of the most-used open codebases in robot learning.
That matters because the money in robotics AI is flowing to closed labs. Figure AI was valued at $39 billion in its September 2025 Series C and Skild AI at $14 billion in January 2026, according to the companies. Hugging Face is betting that a shared, open layer will sit underneath those labs and their rivals. The sections below explain what it has built, what it charges for, and where the bet could fail.
A software company that bought a robot maker to get a body
Hugging Face runs the largest public hub for AI models and datasets. It started LeRobot in January 2024, when the GitHub repository was created, and released it under the Apache 2.0 licence. On April 14, 2025 it announced the acquisition of Pollen Robotics, the Bordeaux maker of the open-source Reachy humanoids founded in 2016, calling it its fifth acquisition and saying it would sell Reachy 2 at $70,000 as its first robot. Terms were not disclosed.
Three months later, on July 9, 2025, Hugging Face and Pollen opened orders for Reachy Mini, an expressive desktop robot 28 cm tall and 1.5 kg, sold as a kit. The blog lists the Lite version at $399 and the wireless version, with a Raspberry Pi 4, battery and accelerometer, at $499, both plus tax and shipping, with delivery in about 90 days and a warning that the product is in an early development phase.
So the company now spans a $399 kit and a $70,000 research humanoid. The question is what ties them together.
LeRobot is the product, and it now covers the whole learning loop
LeRobot v0.6.0, published on July 7, 2026, shows the scope. It added three world-model policies (VLA-JEPA, LingBot-VA, FastWAM), new vision-language-action models including NVIDIA's GR00T N1.7 and the Allen Institute for AI's MolmoAct2, a reward-model API, six simulation benchmarks under one evaluation command, and a lerobot-rollout tool that records human corrections during deployment in DAgger style so failures become training data. Training can run on Hugging Face's paid HF Jobs compute. NVIDIA said a day earlier that it is bringing GR00T, Isaac Teleop, Isaac Lab-Arena and Jetson Thor support into LeRobot. Version 0.6.1 followed on August 3.
LeRobot documents the low-cost SO-100 and SO-101 arms as supported hardware; the v0.6 notes show MolmoAct2 running zero-shot on them. In June 2025 Hugging Face also released SmolVLA, a compact VLA trained on community data recorded with LeRobot.
- LeRobot
- Open-source library (Apache 2.0)
- Repo created Jan 26, 2024
- Free; optional paid HF Jobs compute
- Pollen Robotics
- Acquisition (Bordeaux, founded 2016)
- Apr 14, 2025
- Terms not disclosed
- Reachy 2
- Mobile research humanoid
- Sold by Hugging Face from Apr 2025
- $70,000
- SmolVLA
- Compact VLA trained on LeRobot community data
- Jun 3, 2025
- Open model
- Reachy Mini Lite / wireless
- Desktop robot kit, 28 cm, 1.5 kg
- Orders opened Jul 9, 2025
- $399 / $499 plus tax and shipping
- LeRobot v0.6.0
- World models, VLAs, reward models, 6 benchmarks
- Jul 7, 2026
- Free
Sources: Hugging Face blog posts (Pollen acquisition, Reachy Mini, LeRobot v0.6.0), GitHub API. Prices as published by the company.
As of Oct 16, 2026
The pattern is that every model, from NVIDIA, AI2 or academic labs, arrives as a LeRobot policy that anyone can fine-tune on their own dataset and push back to the Hub. That leaves the commercial question open.
MEKAWIRE analysis
Hugging Face is building the GitHub of robot learning, not a robot company: its leverage is where data and policies are exchanged, and the hardware is a customer-acquisition cost.
The evidence is in the price ladder. Reachy Mini Lite costs 0.57% of Reachy 2 ($399 divided by $70,000) and 6.8% of Unitree's R1 humanoid at its $5,900 international launch price. No robotics revenue has been disclosed, and MEKAWIRE holds no verified funding round for Hugging Face, so the robotics unit cannot be valued on its own. What can be measured is adoption: LeRobot has gathered about 860 GitHub stars a month since January 2024 (28,209 stars over 32.7 months), and five tagged releases shipped between February 27 and August 3, 2026.
- Hugging Face / Pollen
- Reachy Mini Lite
- $399
- Jul 2025
- Company price
- Unitree
- R1 humanoid (international)
- $5,900
- Jul 2025
- Company price
- Hugging Face / Pollen
- Reachy 2
- $70,000
- Apr 2025
- Company price
- Physical Intelligence
- Series B post-money
- $5.6bn
- Nov 2025
- Reported
- Skild AI
- Series C post-money
- $14bn
- Jan 2026
- Company announcement
- Figure AI
- Series C post-money
- $39bn
- Sep 2025
- Company announcement
From MEKAWIRE verified rows. Reachy Mini Lite share of Reachy 2 price = 399 / 70,000 = 0.57%; of Unitree R1 = 399 / 5,900 = 6.8%. LeRobot stars per month = 28,209 / 32.7 months since Jan 26, 2024. Hugging Face valuation not in MEKAWIRE verified data.
As of Oct 16, 2026
The strongest counter-argument is that open libraries rarely capture the value they create. Physical Intelligence has published its own open pi0 and pi0.5 weights, and NVIDIA ships GR00T through its own Isaac tools, so the model owners can host elsewhere. Research shows the dependency too: an October 15 preprint from Shanghai Jiao Tong University built its real-world reinforcement-learning system on Physical Intelligence's open pi0.5, not on a Hugging Face model.
Bull case: LeRobot becomes the default format for robot datasets, as the Hub became for language models, and Hugging Face earns on compute (HF Jobs), enterprise hosting and hardware. Bear case: the large labs keep their best data private, Chinese stacks consolidate around domestic tools, and LeRobot stays a teaching tool while the revenue sits with model and robot makers.
Three signals will settle it: whether a commercial robot maker ships a product whose policy is trained and distributed through LeRobot in 2027; whether NVIDIA's Cosmos 3 integration, promised as coming soon in July, lands in a 2026 release; and whether Hugging Face discloses robotics revenue or Reachy Mini delivery volumes in its next funding announcement.
- 28,209 GitHub stars and 5,893 forks on Oct 16, 2026 (GitHub API)
- Open-source robot-learning library LeRobot (Apache 2.0, repo created Jan 26, 2024) plus Pollen Robotics hardware
- Announced Apr 14, 2025; fifth acquisition; terms not disclosed
- GR00T, Isaac Teleop, Isaac Lab-Arena and Jetson Thor support coming into LeRobot (NVIDIA, Jul 6, 2026)
- LeRobot v0.6.0 on Jul 7, 2026 (world models, GR00T N1.7, MolmoAct2, reward models, six benchmarks); v0.6.1 on Aug 3
- Reachy Mini Lite $399, wireless $499 (plus tax and shipping); Reachy 2 $70,000
- Not disclosed
Why it matters
Robot learning is splitting into closed vertical stacks (Figure, Tesla, 1X) and an open layer where models, datasets and evaluation are shared. Hugging Face is the most-used home for that open layer, so its choices about formats, licences and hardware shape what universities, start-ups and integrators can build without a licence from a large lab.
For investors the relevance is indirect: Hugging Face does not report robotics revenue, but the tools that become standard in LeRobot decide which model vendors (NVIDIA, AI2, Physical Intelligence) reach developers fastest.
Rival analysis
NVIDIA is partner and potential rival: it supplies GR00T and Isaac tools into LeRobot but also runs its own developer channel. Physical Intelligence open-sources weights through its own openpi repository. Google DeepMind offers a Gemini Robotics SDK to trusted testers only. In China, AgiBot and others publish datasets and models on their own platforms.
On hardware, Unitree's R1 at $5,900 and low-cost arms from many Chinese suppliers compete for the same student and lab budgets that Reachy Mini and SO-101 target.
Valuation context
MEKAWIRE holds no verified Hugging Face funding row, so we do not state a valuation. Peers that monetise models directly are priced at $5.6bn (Physical Intelligence, reported), $14bn (Skild AI) and $39bn (Figure AI). Hugging Face's robotics unit should be viewed as an option on platform revenue (compute, enterprise hosting) rather than on model licensing.
The Pollen purchase price was not disclosed, so no return on that acquisition can be computed.
Supply-chain implications
Reachy Mini uses commodity parts, a Raspberry Pi 4 in the wireless version, a wide-angle camera, four microphones and a 5 W speaker, sold as a kit to cut assembly cost. Reachy 2 is built by Pollen in France. LeRobot's NVIDIA integration ties high-end deployment to Jetson Thor modules, while the SO-100/101 arms rely on low-cost serial servos widely sourced from China.
Signals to watch
1) Whether Cosmos 3 support, promised by NVIDIA in July as coming soon, ships in a LeRobot release before year-end 2026. 2) Any disclosure of Reachy Mini delivery numbers or robotics revenue by Hugging Face. 3) A commercial robot product in 2027 whose policy is trained and distributed through LeRobot.
Analyst view
Thesis: Hugging Face is building distribution for robot learning, not robots. Confidence: medium. Adoption data (stars, forks, release cadence, NVIDIA integration) supports the platform role; the lack of disclosed revenue and the ease with which model owners can host elsewhere limit how much value it captures.
Questions you should be asking
What share of HF Jobs compute is robotics training? How many Reachy Mini units have shipped against orders? Will LeRobot dataset formats be offered for standardisation to industry bodies? Is there a plan for a commercial support tier for factories and integrators?