Robotics AI company Physical Intelligence said on 13 October 2026 that it is “making our latest models available for training and inference through the Pi API beta.” Access is by application: companies ask to join through a waitlist, so this is not a public launch. The company has not announced pricing.
What the API does
Physical Intelligence builds what are called robotic foundation models: large AI models trained to give robots general physical skills, which can then be adapted to specific jobs. Adapting them normally needs machine learning expertise, computing power and data pipelines.
The API is meant to take that work off developers’ hands. In its blog post, the company explains: “You upload data, launch fine-tuning on compute and infrastructure we manage, and access fine-tuned models via a hosted gateway to run inference.”
In plain terms, a robot builder records examples of the task, sends them to Physical Intelligence, has the model trained on that data on the company’s computers, and then runs the trained model over the internet to control the robot. “Inference” means running the model to make decisions.
The company says the model currently offered through the API “already provides more data-efficient fine-tuning and improved capability for our partners,” and that it plans to keep giving API users its latest models.
The company’s examples
Physical Intelligence says it gave “a small number of companies” early access. It highlighted two examples, which are its own showcase results rather than independent tests:
- Actor, which makes autonomous heavy machinery. The company says Actor’s self-driving wheel loader scoops material and loads it into a dump truck “during a real commercial deployment without any operator inside the machine.”
- Almond, which builds a two-armed robot called Axol. According to Physical Intelligence, Axol now deburrs (smooths the rough edges of) spacers used to build Axol itself, “running a model the team obtained after iterating with the Pi API for four days.”
The company says most of its showcased applications, “from Japan to Switzerland to deployments across the United States,” ran with cloud-hosted inference through the API. It also admits that the early deployments showed areas to improve, “including network connectivity challenges for some.” Running a robot’s brain in the cloud means a weak internet connection can affect how it performs.
How to join
The company says anyone who wants to take part should “tell us what you want to build” through its waitlist. It has not said how many companies will be admitted, when wider access might follow, or what it will cost.
What this means
- Robot AI is being sold as a service. Builders can rent a trained robot brain instead of developing their own.
- Access is limited. The beta is by application, and pricing has not been announced.
- The results are the company’s own examples. No independent evaluation of the API has been published.
Sources
- Physical Intelligence, “Pi API beta” blog post, 13 October 2026. https://www.pi.website/blog/api-beta
Related stories
- NEURA Robotics Unveils Neuron AI, a Model That Lets Robots Combine Sight, Touch, Sound and Memory
- Odyssey Opens Its Odyssey-3 World Model to the Public; Its Headline Physics Score Is a Best-of-Eight, and Every Benchmark Is Odyssey’s Own
- NVIDIA’s “Physical AI” Play: Teaching Robots to Think, Talk, and Actually DO Things
