Physical Intelligence opens its π robot models to outside developers with the Pi API beta

What did Physical Intelligence launch?

Physical Intelligence, the robotics lab behind the π0 and π0.5 robot foundation models, announced the Pi API beta on 13 October 2026. The service makes the company’s latest robot model available for training and inference by outside developers, rather than only through the lab’s own research partnerships.

The pitch is simple. General-purpose robot models such as π0 and π0.5 already carry a lot of physical knowledge, the company says, which cuts the data and effort needed to adapt them to a new task, environment or robot. But adapting them still takes machine-learning infrastructure, compute and data pipelines. The Pi API is meant to take that work off the developer’s plate.

Physical Intelligence says the latest model offered through the API already fine-tunes more data-efficiently and performs better for its partners than earlier versions. It does not name or benchmark that model in the announcement. The company also says it can update the models behind the API quickly and plans to keep offering its newest ones there.

How does the Pi API work?

The workflow has three steps. A team uploads its own robot data, usually demonstrations collected by teleoperating the robot. It then launches fine-tuning of a π base model on compute and infrastructure that Physical Intelligence manages. Finally it calls the fine-tuned model through a hosted gateway. In hosted inference, the robot sends its observations to the model running on Physical Intelligence’s servers and gets actions back, instead of running the model on its own computer.

This is the same loop robotics teams already run by hand: collect demonstrations, train a policy, test it on the robot, and use the failures to decide what data to collect next. Physical Intelligence says it built the API so developers can run that loop themselves, without waiting for the lab’s researchers. If the idea of adapting a big pretrained model to your own examples is new, our explainer on fine-tuning covers the basics.

In practice, a π model in this setting is a vision-language-action model. It looks at camera images, reads a text instruction, and outputs motor commands for the robot’s arms, grippers or wheels. Fine-tuning teaches it the specifics of one robot body and one job, such as how hard to squeeze a strawberry or where a part should sit on a fixture.

What did early users build with it?

Physical Intelligence published short case notes and videos from two groups: partners it has worked with closely, and a small set of companies given early access to try the API on their own. Most of the deployments shown ran with cloud-hosted inference, in places ranging from Japan to Switzerland and across the United States. Highlights from the company’s descriptions:

• Argon Robotics, which builds kitchen robots, had a fine-tuned model racking dishes within 35 hours of getting API access and switched its deployment to it the next day.
• Chef Robotics trained its first two-arm model for burger assembly from 33 bimanual demonstrations collected the day before, and had it running on the robot the next day without running training itself.
• Deft, which builds wheeled humanoids for car plants, says it more than doubled its success rate on a part-loading task in six days, training a new model overnight each time.
• Foundry uploaded about 500 demonstrations of an engine-assembly step and was testing hosted, fine-tuned models on its factory floor the next day.
• Frontier Machines trained its own packaging model within a week that matched a policy Physical Intelligence’s researchers had earlier tuned for it, then deployed it at a customer site.
• General Instinct says its server-assembly policy, trained within three days, matched an existing fine-tuned open-source model while using 5 per cent of that model’s data.
• Omakase Robotics says its first model poured tea 10 times out of 10 from 97 demonstrations, and ran on its other D1 humanoids without the per-robot camera calibration its earlier policy needed.
• Almond’s dual-arm Axol robot deburrs spacers used in its own assembly, using a model the team got after four days of iterating with the API.

Which industries are already using π models?

The partner list is broad. Actor runs an autonomous wheel loader in Utah that scoops material into a dump truck with no operator in the cab. AgRobotics picks strawberries with a three-fingered gripper from a robot that drives in the furrow like a field crew. CloudChef portions ingredients in commercial kitchens, and an unnamed dairy partner uses an autonomous tractor to push feed towards cows for hours at a time.

In warehouses, Ultra packs e-commerce orders and Yondu picks items from bins into totes. Yondu’s own software navigates and checks each bin, then calls a π model through the API for the pick, with teleoperators correcting it when needed. Formic, which says it automates more than 150 US manufacturers, brought π models to three robot types working from the API documentation. Weave’s laundry-folding robots work in homes, hotels, shops and laundromats. Intuition.dev says it trains and deploys in under five hours, and robot.com uses a humanoid to load orders into sidewalk delivery robots.

The range matters because robot foundation models are still trying to prove they generalise across very different bodies. Nature Will’s Natura-1, announced the same day, attacks the problem from the data side by learning hand control from human video. Physical Intelligence is betting instead on a shared model that each customer specialises with its own robot data.

Who can get access, and what does it cost?

Access is through a waitlist form. Physical Intelligence says it is starting small and onboarding teams in batches. The form asks how many engineers work on robotics, who would train the policy, and whether they have already integrated a model into a robot’s control stack, collected training data, fine-tuned a policy or compared policies such as ACT, diffusion policy and π0.5.

It also asks how many hours of teleoperation data a team has, how many hours of simulation data it plans to use, how many robots it has, what the robot bodies are, and whether inference needs to run in the cloud, locally on the team’s own hardware, or both. The announcement itself only describes cloud-hosted inference, so the local option on the form should be read as a question, not a promise.

Neither the blog post nor the waitlist page lists prices, usage limits, supported data formats, latency figures or a date for general availability. Teams should not plan budgets around the API until those terms are published.

How is this different from the open-source openpi models?

Physical Intelligence already publishes open models. Its Apache 2.0 openpi repository holds base checkpoints for π0, π0-FAST and π0.5, pre-trained on more than 10,000 hours of robot data, plus fine-tuned checkpoints for platforms such as ALOHA and DROID and code to fine-tune them yourself.

Running those yourself has real requirements. The repository lists more than 8 GB of GPU memory for inference, more than 22.5 GB for fine-tuning with LoRA, and more than 70 GB for full fine-tuning on an A100 or H100. It is tested only on Ubuntu 22.04, and its training script does not support multi-node training.

The API is a different trade. Teams give up running the model on their own terms in exchange for the company’s newest model, managed training and hosted serving. The openpi README also warns that π0 was built for the lab’s own robots and that not every attempt to adapt it to other platforms will succeed; the API’s partner stories are the company’s answer to that worry.

Why a hosted API matters for robotics

For language models, renting a fine-tuned model behind an API is routine. For robots it is newer, because the model has to drive physical hardware in real time and the data is expensive to collect. Physical Intelligence argues that every robotics application, from construction to surgery to warehouses, is hard enough to need its own company, and that those companies should not each have to solve general physical intelligence from scratch.

There is also a feedback loop for the lab. The company says that seeing where its models succeed and fail in real deployments helps it make the next generation more broadly useful. In AI Lookout’s reading, the API turns Physical Intelligence from a research lab with a handful of partners into a platform whose customers keep showing it new tasks.

The speed claims are the most striking part. If a team really can go from fresh demonstrations to a working model overnight, the bottleneck in robot deployment shifts from training infrastructure to collecting good data and testing on real machines.

What are the limitations and open questions?

Every result in the announcement is reported by Physical Intelligence or its partners, mostly as short descriptions and videos. There is no shared benchmark, few success rates, and no independent test. Phrases such as “high task success” are not numbers, and the strongest claims, such as matching an open-source model with 5 per cent of the data, cannot be checked from the post.

Cloud inference is a real constraint for machines that move. Physical Intelligence itself says the early deployments highlighted areas for improvement, including network connectivity problems for some teams. A robot that depends on a remote model needs a plan for lost or slow connections.

Model updates cut both ways. The company plans to swap in newer models quickly, which may improve results, but teams that validate a policy for a customer site will want to know how versions are pinned and how long older ones stay available. The announcement does not say, and it does not spell out how uploaded robot data may be used. Teams should read the beta terms closely before sending production data.