Developers can now fine-tune Physical Intelligence's robot models via hosted API
The company behind π0 and π0.5 is now giving external developers access to the models for fine-tuning and inference via a hosted API. Early partners describe an autonomous wheel loader in commercial operation and a deburring model built in four days — but all performance figures come from Pi itself.
Physical Intelligence (Pi) announced on October 13, 2026 that it is opening an API beta giving external developers access to the company's latest robot foundation models, π0 and π0.5, for both training and inference. The announcement, published on Pi's own blog (pi.website/blog/api-beta), is, according to the company, the first step toward offering a shared "physical intelligence" platform that other companies can build on.
In short: Pi wants to do for robots what API-delivered language models did for software — turn an advanced model into something a developer can adapt and call from their own code, without needing in-house model expertise or compute infrastructure.
How the Workflow Works
The product description in the announcement is concrete. Developers upload their own data, kick off fine-tuning on compute and infrastructure that Pi manages, and then access the fine-tuned model through a hosted gateway to run inference. In other words: the customer does not need their own training cluster or serving stack — the entire loop from data to deployed model runs through Pi's systems.
Pi writes that the model currently served through the API "already delivers more data-efficient fine-tuning and improved capacity for our partners," and that the company plans to keep updating the models behind the API, since Pi can update them quickly. That means customers should, in principle, get access to Pi's newest and most capable models as they arrive — model versioning along the lines of what we know from hosted language model APIs, though the beta does not yet document the details.
Why Foundation Models, and What Is Still Missing
Pi's argument for the product rests on its home turf of general-purpose robot foundation models. According to the company, such models provide a strong foundation of physical knowledge and skills that "substantially reduces the data and effort needed to adapt them to new tasks, environments, or robots." But Pi is also clear about the limitation: adaptation still requires ML infrastructure, compute, and complex data pipelines, which the company itself describes as "difficult in their own right."
This is precisely what the new workflow aims to remove. So the foundation model does not drive the cost of adaptation to zero — it shifts the cost from the customer to Pi, as a service. The announcement thus positions the fine-tuning and inference work itself as the product, not just the model weights.
Worth noting: Pi says nothing in the source material about pricing, availability timelines, rate limits, or which robot types and hardware platforms the API supports. For a beta of this scope, those are open questions that come with it.
The Early Partners
Pi describes how a small number of companies were given early access to the API to test whether they could improve their autonomy on their own. The company writes that it was "impressed by how quickly they got started and began making progress." The list of partner names given in the announcement includes Actor, Almond, Argon Robotics, Chef Robotics, Deft, Dream Machines, Formic Robotics, Foundry, and Frontier Machine — with the caveat that part of the list comes from truncated UI text in the partner carousel on the page, so the full roster cannot be confirmed precisely from the available material.
Two partner examples are highlighted:
Actor deploys heavy machinery autonomously. According to Pi's description, their autonomous wheel loader scooped material and loaded it into a dump truck "in a real commercial deployment with no operator inside the machine." If so, this is an example of a robot in paid operation, not just in a laboratory.
Almond builds Axol, a dual-armed robot aimed at developers in "physical AI." The goal, according to the company, is for Axol to be able to build itself. In Pi's example, Axol performs deburring — the removal of burrs and edges — on spacers used in the robot's own assembly, with a model the team reportedly achieved after four days of iteration via the Pi API.
Both examples are Pi's own descriptions of its partners' results. There is no independent verification of either the four-day iteration or the operator-free operation in the available material.
The Iteration Loop as a Design Principle
Perhaps the most interesting part of the announcement is how Pi justifies the API from its own collaboration experience. Through the early partnerships, the company saw that the process for improving performance was "roughly the same" across projects: collect demonstrations, train a model, test it on the robot, and use the errors to decide which data should be collected next time.
Pi writes that the API is explicitly built to let developers run this iteration loop themselves, with the company's latest models. In other words, the product is not "a model" but the loop itself: demonstrations in, training on managed compute, testing on the robot, new data collection, and so on. It is workflow thinking drawn from machine learning — where progress is measured by how fast you can cycle through the loop — packaged as a developer service.
If this holds up in practice, that is where the "API moment" for robotics lies: not in robots suddenly becoming generally capable, but in the time and expertise needed to adapt a model to a concrete task being reduced to something a developer team can handle.
Uncertainty and Open Questions
Several factors should be kept separate from the news value of the launch:
- All performance claims are Pi's own. That fine-tuning is "more data-efficient," that Actor runs in commercial operation without an operator, and that Almond's model was developed in four days all stem from Pi's own announcement and are not independently verified. No third-party reporting or benchmarks are available in the material at hand.
- The beta's details are undocumented. Pricing, when the API will become generally available, any rate limits, supported hardware, and how model versioning is handled in practice are not stated.
- The partner list is partly uncertain. The list of names comes partly from truncated UI text, so the complete list of early users cannot be determined precisely.
- What "latest models" means in practice. Pi says the API serves the company's latest models and will keep updating them, but does not specify which version levels of π0 and π0.5 apply, or how updates affect already fine-tuned customer models.
What can be established as fact is that Physical Intelligence on October 13, 2026 launched an API beta with external access to its robot foundation models for training and inference, with a workflow in which the customer uploads data, fine-tunes on Pi's infrastructure, and runs inference through a hosted gateway — and that the company itself presents a series of partnerships as proof of concept. How well the concept works outside Pi's own presentation is a question the next round of reporting will have to address.