Hugging Face's New Robotics Model Runs on a Macbook

The Growing Accessibility of Advanced Robotics

The creation of complex robotics projects is becoming increasingly achievable for hobbyists and developers working from home.

This week saw the release of SmolVLA, an open AI model designed for robotics, by the AI development platform Hugging Face. The model was trained using datasets shared by the community under compatible licenses.

SmolVLA's Performance and Goals

Hugging Face asserts that SmolVLA surpasses the performance of considerably larger models in both simulated and real-world robotics applications.

According to a blog post by Hugging Face, SmolVLA’s primary objective is to broaden access to vision-language-action (VLA) models and to expedite research focused on creating versatile robotic agents.

The platform emphasizes that SmolVLA isn't just a compact, yet effective model; it also represents a methodology for both training and evaluating generalist robotics technologies.

Hugging Face's Expanding Robotics Ecosystem

SmolVLA is a key component of Hugging Face’s ongoing initiative to build a comprehensive ecosystem of affordable robotics hardware and software.

Last year, the company introduced LeRobot, a resource encompassing robotics-specific models, datasets, and tools. More recently, Hugging Face completed the acquisition of Pollen Robotics, a French robotics startup, and subsequently launched a range of cost-effective robotics systems, including humanoid robots, available for purchase.

Technical Specifications of SmolVLA

SmolVLA comprises 450 million parameters. These parameters, often referred to as “weights,” are the internal elements of the model that dictate its operational behavior.

The model was trained utilizing data sourced from LeRobot Community Datasets and robotics datasets specifically shared on Hugging Face’s AI development platform.

Hardware and Software Compatibility

Hugging Face states that SmolVLA’s relatively small size allows it to operate on a single consumer-grade GPU – or even a MacBook.

Testing and deployment are possible on reasonably priced hardware, including the company’s own robotics systems.

Asynchronous Inference for Faster Response

Interestingly, SmolVLA incorporates an “asynchronous inference stack.” Hugging Face explains that this feature enables the model to decouple the processing of a robot’s actions from the processing of its sensory input (sight and sound).

This separation allows robots to react more swiftly in dynamic and rapidly changing environments.

Availability and Early Adoption

SmolVLA is currently available for download from Hugging Face. A user on X has already reported successfully utilizing the model to control a robotic arm from a different manufacturer.

The Broader Landscape of Open Robotics

It is important to note that Hugging Face is not alone in the emerging field of open robotics.

Companies like Nvidia offer a suite of tools for open robotics, and K-Scale Labs is developing the building blocks for what they term “open-source humanoids.”

Other significant players in this sector include Dyna Robotics, Physical Intelligence (backed by Jeff Bezos), and RLWRLD.