Hugging Face Founder Thomas Wolf: Training Robots Should Be as Simple as Writing Applications
Hugging Face co-founder Thomas Wolf stated in an interview with Bloomberg that he hopes training robots will be as simple as writing code and creating applications, allowing everyone to own and modify their own robots.
The company launched the open-source bipedal robot Microduck, priced at around $399 to $400, developed by the French company Pollen Robotics, which was acquired last year. This is the second consumer-oriented robot, following the Reachy Mini. The robot is approximately 25 cm tall and weighs less than 800 grams. It can walk, sit, kick, roller skate, grasp objects with its beak, and recover balance after common falls. It comes with a game controller for users to play before coding. Developers can teach new skills using reinforcement learning, simulation, and open-source tools like Python and JavaScript, with strategies running at about 50 Hz on the device and uploaded to the Hugging Face model library. Wolf mentioned on Bloomberg's tech program that during the initial sales period, there was an order every four seconds; the company later reported selling over 10,000 units within a week, generating over $5 million in sales, with deliveries delayed past the originally planned Christmas timeline. He declined to comment on rumors of NVIDIA's $12.9 billion acquisition, stating that the company often receives acquisition and investment inquiries.
The low-cost open-source approach turns robots into customizable toys, but this does not equate to meeting household capabilities. There is still a gap between simulation and real machines. The shipping delays indicate that the supply chain cannot keep up with the promotional pace.
Families and enthusiasts are grouped under the same SKU. The model library treats behaviors as downloadable applications.
In market mechanics, buyers include developers, enthusiast families, and educational users, while the seller is Hugging Face, which sells hardware and community models. The driving force is the product and interviews, categorized as a consumer hardware event. Funds flow from pre-orders into complete machines and chip inventory. Beneficiaries include open-source robot kits and simulation tools, while those under pressure are high-priced humanoid projects that close ecosystems and those trading duck sales as a universal humanoid turning point.
Source: Public Information
ABAB AI Insight
Thomas Wolf applies the model marketplace logic of Hugging Face to joints: behaviors are files, and the duck is the runtime. The $400 price and controller lower the barrier of "first learn to walk, then talk intelligence." The acquisition of Pollen turns the French lab into shelf goods. An order every four seconds and Christmas delays are two sides of the same coin: the open-source community will place orders, but factories will not ship to the rhythm of tweets. The Reachy Mini has validated that desktop robots can sell at the scale of 100,000 units, and Microduck has transformed its form into a duck more suitable for short videos. The NVIDIA acquisition rumors place open-source neutrality opposite valuation, and Wolf's refusal to comment keeps the options on the table. Training like writing applications assumes the simulator is close enough to the floor.
The capital path is hardware gross margins plus hub network effects. Selling ducks aims to bring more strategies back to the website. The motivation is to democratize embodied intelligence; the strategy is to occupy the developer desktop with low-cost SKUs rather than first creating factory workers. Chinese chip solutions are entering open-source American and French brands, and supply chain politics will emerge alongside shipping orders.
Analogous objects include Raspberry Pi, LEGO Mindstorms, and Boston Dynamics treating videos as products. Consumer-grade open-source bipedal robots sit between toys and teaching tools. The phase is shifting from model hosting to entity hosting.
This is a prelude to technological substitution. The alternative is that "robots can only be flashed by large manufacturers." The mechanism is: open-source strategies combined with purchasable joints shift skill distribution from firmware to app stores. Until the substitution is complete, recovering from falls and roller skating remain demonstrations, not household tasks. Whoever controls the distribution of behavior files will set the price for the next duck.
ABAB News · Cognitive Laws
- When training can be as simple as writing applications, robots will start to resemble personal computers.
- An order every four seconds proves there is demand, while Christmas delays prove there is production capability.
- Open-source entities turn skills into downloadable files, making factories the only bottleneck.