Hugging Face believes its latest robotics model is efficient enough for a MacBook.
Hugging Face believes its latest robotics model is efficient enough for a MacBook.
Building advanced robotics projects at home is becoming simpler.
Hugging Face published SmolVLA, an open AI model for robots, this week. Hugging Face says that SmolVLA beats bigger robotics models in virtual and real-world contexts using “compatibly licensed,” community-shared datasets.
Hugging Face says in a blog post that SmolVLA seeks to democratize VLA models and expedite generalist robotic agent research. SmolVLA is a lightweight, competent model for teaching and assessing generalist robotics systems.
Hugging Face is fast building a low-cost robotics hardware and software ecosystem, including SmolVLA. The business released LeRobot last year, a bundle of robotics models, datasets, and tools. Hugging Face recently bought Pollen Robotics, a French robotics business, and released many cheap robotics systems, including humanoids.
Trained using LeRobot Community Datasets, specifically annotated robotics datasets provided on Hugging Face’s AI research platform, SmolVLA has 450 million parameters. The fundamental components of a model, called parameters or “weights,” govern its behavior.
Hugging Face says that SmolVLA can be tested and implemented on “affordable” technology, including its robotics systems, and runs on a single consumer GPU or MacBook.
In an intriguing twist, SmolVLA enables a “asynchronous inference stack,” which Hugging Face claims lets the model differentiate robot behaviors from what it sees and hears. The business writes on its blog that “[b]ecause of this separation, robots can respond more quickly in fast-changing environments.”
