Physical Intelligence API waitlist (how access works)

Access to the Physical Intelligence pi API runs through one page: the waitlist form on the company's own site, checked 26 October 2026. You fill in one application about your team, your robot, and your data. The company reads it and admits teams in batches. There is no self-serve signup, no published price, no published rate limit, and no API documentation page. Each of those absences was checked today by asking the domain directly.

This page is the walkthrough the search results do not have. The waitlist form itself explains nothing; it just asks. News coverage from the launch day stops at "there is a waitlist." This page gives you the whole process. You get every question the form asks, quoted from the live page. You get what each question screens for, what admission gets you, and what to do while you wait. Every fact carries the date it was read.

What the pi API is

Physical Intelligence is a robotics model company. It builds general-purpose policies, a model family it writes as pi 0, pi 0.5, and newer. A policy takes camera images and a text instruction, and outputs the actions a robot arm should run. On 13 October 2026 the company announced the Pi API beta. The hosted service has three steps. You upload your robot demo data. You fine-tune one of their models on compute they manage. You run the resulting policy on your robot through hosted inference. The post's own summary of the loop: "Upload your data, fine-tune a pi model, and run it on your robot through hosted inference. Use evaluation data to guide the next iteration."

The beta is the company's first external product surface. Before it, the public artifacts were research posts and the open weights repository described below. The post ends with one access instruction: if you would like to join the beta, tell them what you want to build. That line points at the waitlist.

Three pi APIs, one search box

Before the mechanics, a warning the search results make necessary. Search for pi website api apply and the robot company's waitlist page ranks first. Five of the seven other rows on page one belong to an unrelated product: the PI System from OSIsoft. That is industrial software for process data. Its PI Web API has its own docs, courses, and tutorials. The remaining two rows are pi.dev, which is not this company either, and a Raspberry Pi forum thread. Search for pi api and you may also meet Inflection's Pi chatbot, or Raspberry Pi computers. This page is about none of those. It is about the robotics company at pi.website. One more naming note from today's read: the company also owns the physicalintelligence.company address, and it redirects to pi.website.

The waitlist, step by step

The application lives at one address, the waitlist page. Its own page title is "API access". The flow has three moves.

First, you fill the form. Every field is on one page; there is no account creation first. Second, the company reviews. The form's opening lines, quoted verbatim from the page read on 26 October 2026: "We're starting small and onboarding in batches. Tell us what you're building, and we'll reach out soon." Third, if admitted, the company contacts you. The post gives no other admission path. There is no dashboard to watch and no public status page for the queue.

That quoted sentence is the entire published account of the timeline. Two facts sit inside it. Access arrives in batches, so the queue moves when a batch opens, not when you apply. And the company reaches out to you. After you submit, the ball is in their court. No page on the site publishes a batch size, a batch cadence, or an average wait.

Every question on the application

The form is the real gate, so here is the complete field list, quoted from the live page on 26 October 2026. Field labels are in quotes; helper text below each field is described. The Greek letter the publisher writes as its model prefix is rendered here as pi.

About you and your team. Six fields: "Name", "Work email", "Company or organization", "Company website", "HQ location" (helper: city, country), and "Team size". The team size helper reads "Engineers and researchers working on robotics". That question wants a number. The helper narrows it to people who work on robotics, not headcount.

Who trains the policy. A free-text block headed "Who on your team would be training this policy?" with the helper "Include their role." This is a person question, not a tools question.

Hands-on work. A checkbox set headed "Which of these have they personally done?" with the instruction to select all that apply. The options, in order: "Integrated a model or external API into a physical robot's control stack". "Collected and prepared robot data for training". "Trained or fine-tuned a policy and run it on a physical robot". "Compared policies (e.g. ACT, diffusion policy, pi 0.5) through repeated robot evaluations". "Debugged a policy". And "None of these yet". It measures not credentials but one thing: whether your team has stood in the exact loop the API automates.

The task. "What specific task do you want to accomplish with the pi API?" The helper text is unusually useful. It reads: "A few sentences are enough. A link to an existing video, results page, or repository is helpful but optional."

Your data. Two number fields. "How many hours of tele-op data have you collected?" And: "How many hours of simulation data do you plan to train on?" The first asks what you already have. The second asks what you intend to build.

Your robot setup. "Number of robots available", then "Describe your primary robot embodiment(s)". The helper: "Please be very descriptive if it's not an off-the-shelf provider." Then a four-way choice, "Where do you need to run inference?" The options: "Cloud", "Local (on your own hardware)", "Both", or "Not sure yet". The local option exists even though the post's examples mostly ran cloud-hosted.

Success and intent. Two more fields. "Success criteria for this task". And a dream-task prompt: "If pi could do any task for you, what would it be?" The second reads like product research. The page marks fields required unless noted.

The referral. The last question before the submit button: "How did you hear about this waitlist?" The helper: "Please name the person who referred you or where you found the link." Then one button: "Submit application".

Count the fields and a pattern appears. The form asks about your data hours and robot count before anything that resembles billing. It never asks about money. There is no payment field, no budget field, and no plan choice on the page.

What the questions signal

No public document explains how applications are screened, so read the form itself. It asks for three things a robotics team cannot fake quickly: trained people, collected data, and working robots. Every question maps onto one step of the fine-tune loop the post describes. A team that already collects demos and tests policies can onboard without hand-holding. That is the plain reading of the form. It is a reading, not a published rule.

The form does leave one door open. The checkbox "None of these yet" exists, so a beginner can apply honestly. But the data and robot fields still expect numbers. A credible form without an operating robot program is hard to write. The referral field is the other signal. A form that asks you to name your referrer expects introductions to circulate.

What admission gets you

The post says it in its own words. You "launch fine-tuning on compute and infrastructure we manage" and you "access fine-tuned models via a hosted gateway to run inference". Three published properties matter to an applicant.

The model is theirs and it moves. The post says the latest model in the API already gives more data-efficient fine-tuning. It commits to rotation in one line: the company can update the models behind the API rapidly, and plans to keep enabling access to its latest and most capable models. It never names the version. Which checkpoint an admitted team gets is not published. The served model is "the latest", whatever that means on your date.

Inference is hosted by default. Most early work shown in the post ran with cloud-hosted inference. The deployments span Japan, Switzerland, and the United States. The form's local-inference option shows the company at least asks where you need inference to run.

The company flags its own weak point. The post says these early deployments "highlighted areas for improvement, including network connectivity challenges for some". A robot loop that depends on a hosted endpoint depends on the network between them. That caveat is the publisher's own, and any applicant should carry it into planning.

What is not published

Every row below was checked on 26 October 2026. A dash is a fact with a date on it.

Third-party coverage fills some of these gaps with guesses. Some launch-week articles state the API serves pi 0 and pi 0.5, because the post links the open repository for those models. The post itself never says it. Treat any filled-in detail the publisher did not print as the writer's guess. That rule applies to anything on this site too.

Who is already in

The post's own text names two teams. Actor runs autonomous heavy machinery. Its wheel loader scoops and loads in a commercial deployment with no operator inside. Almond builds Axol, a dual-arm robot. Axol deburred spacers used in its own assembly, running a model the team reached after four days with the API. The post also carries logo panels of early teams. Launch coverage counted roughly twenty named partners and early users across manufacturing, farming, and food service. None of that promises anything about your form. It is the published evidence that the batches are real and moving.

While you wait

Two paths exist today, both checked 26 October 2026, and neither needs an invitation.

Run the open weights yourself. The company publishes the openpi repository under Apache-2.0. It holds base checkpoints for pi 0, pi 0 FAST, and pi 0.5, pre-trained on more than 10,000 hours of robot data. The repo's own table sets the hardware bar. Inference wants more than 8 GB of GPU memory; the example card is an RTX 4090. LoRA fine-tuning wants more than 22.5 GB. A full fine-tune wants more than 70 GB: an A100 80 GB or an H100. One command starts a policy server. A small client package talks to it from your robot code. The repo documents Ubuntu 22.04 as the operating system it has been tested on. Its own warning is worth keeping: pi 0 was built for the company's robots, and it may or may not work on yours. The pi 0.5 pricing page on this site holds the deeper treatment of the open-weight path.

Budget a per-call lane here. This platform carries a written-record row for the open pi 0.5 family: the pi05 ids sit in the platform's model data, not on the public list of served models, as of 26 October 2026, with no date attached to that step. Only listed rows render a rate. The table below renders from the platform's model data at read time.

Metering per call is the shape that matters for a robot loop. The pricing units guide translates it into cost per robot hour. The alternatives guide on this site compares every dated option side by side. The host-class report records which major hosts carry robot models at all.

If you apply

The form's own helper text is the advice. Be descriptive about your robot if it is not an off-the-shelf provider; the helper asks for exactly that. Attach a link to a video, results page, or repository with your task answer; the form says it helps. Name your referrer if you have one; the last field asks. Answer the checkboxes honestly, including "None of these yet" if that is the truth. The rest of the form states your data hours and robot count anyway. And write the task in the requested size: a few sentences. Nothing on the published record suggests essays help.

Frequently asked questions

How do I get access to the pi API?

Apply at the waitlist page on the company's site. Fill the one form about your team, your data, and your robot. Then wait for the company to reach out. Admission is by batch; the page says so in its opening lines, quoted above. There is no self-serve signup path and no other public route in, as of 26 October 2026.

What does the pi API application ask?

About twenty fields across grouped sections: your team details, who would train the policy, a hands-on checklist, the task, your data hours, and your robot setup. Then success criteria, a dream-task prompt, and how you heard about the waitlist. The full list, in the publisher's own words, sits in the section above.

How long does the pi API waitlist take?

No one publishes that. The form's line, "We're starting small and onboarding in batches", is the only timeline statement the company has made. No batch size, cadence, or average wait appears on any company page checked on 26 October 2026. Anyone quoting you a wait duration is guessing.

Does the pi API cost anything?

No price is published. The post names no price, the form asks no billing question, and the site has no pricing page today. That is not the same as free. When a price appears, it will be news, and this page's dated record will need it.

Which model does the pi API serve?

The company does not say beyond "the latest". It also reserves the right to change that: the post commits to its latest and most capable models as they update. Launch coverage names pi 0 and pi 0.5 by reading the linked repo. Treat named versions as coverage guesses, not publisher statements.

What can I do while I wait?

Two dated options. Self-host the open weights from the openpi repository; the bar is a GPU with more than 8 GB of memory for inference. Or budget a hosted per-call lane; this platform's pi05 row is a written record not on the public list as of 26 October 2026, with no date attached. The alternatives guide compares every option with its date.

The short version: one form, batches, no price, no published timeline. The form asks who trains, what data you have, and what robot it runs on, and nothing about money. The open weights and the per-call budget path are the two things you can start today. Cost questions about this platform land on the pricing page. You can sign up and create a key without a call from anyone.

Last verified: 2026-10-26