It\u2019s becoming a little easier to build sophisticated robotics projects at home.<\/p>\n
Earlier this week, AI dev platform Hugging Face released an open AI model for robotics called SmolVLA<\/a>. Trained on \u201ccompatibly licensed,\u201d community-shared datasets, SmolVLA outperforms much larger models for robotics in both virtual and real-world environments, Hugging Face claims.<\/p>\n \u201cSmolVLA aims to democratize access to vision-language-action [VLA] models and accelerate research toward generalist robotic agents,\u201d writes Hugging Face in a blog post<\/a>. \u201cSmolVLA is not only a lightweight yet capable model, but also a method for training and evaluating generalist robotics [technologies].\u201d<\/p>\n SmolVLA is a part of Hugging Face\u2019s rapidly expanding effort to establish an ecosystem of low-cost robotics hardware and software. Last year, the company launched LeRobot, a collection of robotics-focused models, datasets, and tools. More recently, Hugging Face acquired\u00a0Pollen Robotics<\/a>, a robotics startup based in France, and unveiled several inexpensive<\/a> robotics systems, including humanoids<\/a>, for purchase.<\/p>\n SmolVLA, which is 450 million parameters in size, was trained on data from LeRobot Community Datasets, specially marked robotics datasets shared on Hugging Face\u2019s AI development platform. Parameters, sometimes referred to as \u201cweights,\u201d are the internal components of a model that guide its behavior.<\/p>\n Hugging Face claims that SmolVLA is small enough to run on a single consumer GPU \u2014 or even a MacBook \u2014 and can be tested and deployed on \u201caffordable\u201d hardware, including the company\u2019s own robotics systems.<\/p>\n In an interesting twist, SmolVLA also supports an \u201casynchronous inference stack,\u201d which Hugging Face says allows the model to separate the processing of a robot\u2019s actions from the processing of what it sees and hears. As the company explains in its blog post, \u201c[b]ecause of this separation, robots can respond more quickly in fast-changing environments.\u201d<\/p>\n SmolVLA is available for download from Hugging Face. Already, a user on X claims to have used the model to control a third-party robotic arm:<\/p>\n \ud83d\ude80 SmolVLA \u2014 feels like a BERT moment for robotics \ud83e\udd16 Big thanks to @RemiCadene<\/a> @danaubakirova @mustash97 @francesco__capu \ud83d\ude4c pic.twitter.com\/TiBkAZGwkM<\/a><\/p>\n \u2014 Xingdong Zuo (@XingdongZ) June 4, 2025<\/a><\/p><\/blockquote>\n It\u2019s worth noting that Hugging Face is far from the only player in the nascent open robotics race. <\/p>\n Nvidia has a collection of tools for open robotics, and startup K-Scale Labs is building the components for what it\u2019s calling \u201copen-source humanoids.\u201d Other formidable firms in the segment include Dyna Robotics, Jeff Bezos-backed Physical Intelligence, and RLWRLD<\/a>.<\/p>\n\n
I tried it on the Koch Arm:
Inference on RTX 2050 (4GB), fine-tuned with just 31 demos, and matches\/outperforms single-task baselines \ud83d\udd25<\/p>\n