Can you run robotics or VLA models on Replicate? The October 2026 answer

No. As of October 26 2026, Replicate hosts no robot control models. Its collections index lists 39 model categories, and none of them is robotics. The collection pages for robotics and vla return not-found errors, while real categories return live pages. The pricing page, the blog, and the changelog carry no robot mention either. The models people hunt for here, OpenVLA, pi 0.5, SmolVLA, GR00T, are open weights you run yourself or get from small new specialists. This page shows the receipts, explains why the gap is structural, and gives you four working paths.

Last verified: 2026-10-26. All amounts on this page are US dollar amounts. Every rate carries its source and its read date, because this market changed twice in 2026 already.

What was checked, and when

An absence claim is worth reading only if it names its evidence and its date. Here is the full census. Every check ran on October 26 2026, against Replicate's own pages, not blog posts about them.

The groups that do exist tell the story by themselves. Image work, video work, speech, music, text, embeddings, 3D shapes, background removal, lip sync, upscaling. The closest neighbors are a vision-models group and an object-detection group. Both take an image in and return an image or a label. Neither takes a robot state in and returns motion commands.

One honesty note. The on-site search box gave no clean zero-result receipt this session. It renders nothing until you click, on every fetch method tried. So this page does not lean on that search. The index, the URL checks with their controls, and the page scans are the receipts, and they are enough. One research record, dated October 25 2026, checked the pricing pages of eleven large hosts. Replicate was one of them. None carried a robotics mention. One day later, the same still holds.

Why the search results look so strange

Search for "replicate robotics models" on Google and you get noise. Five of the first seven rows in the October 26 capture were about machines that copy themselves. A research page on self-replicating AI. The Wikipedia article on self-replicating machines. A Harvard story about living robots that reproduce. One DeepMind robot-model page also ranks, and that is a different company's product. The searcher with a robot arm gets nothing.

Search for "replicate vla model" and a second collision takes over. The top rows belong to VLA-REPLICA, a test rig from the University of Texas at Dallas, published on arXiv in May 2026. It is a cheap, repeatable bench for grading robot policies on real hardware. Any lab can rebuild the same setup. That is what the name means. It has no link to the Replicate platform. It still owns the first three results, and it owns two of the three queries this page targets.

The autocomplete layer shows the same gap from the inside. On October 26 2026, Google's suggestion box for "replicate vla" returns nothing at all. The box for "replicate fl" returns ten flux ideas. The box for "replicate up" returns ten upscaler ideas. The suggestion index is dense for every real Replicate model family. For this one it is silent. One more collision sits nearby: "replicate rob" suggests replika robot, the companion chatbot app.

Even Google's own AI Overview concedes the point. For the query "run vla on replicate", it teaches you to package your own model with Cog, push it to your own account, and call it from there. That is a bring-your-own-container answer. When the best ranking answer to "can I run X there" is "build and upload X yourself", the platform itself has none to give you.

Why Replicate carries no robot models

This absence is not an accident of sorting. It falls out of how the platform bills and how its API is shaped. Both are documented on its own pages, read October 26 2026.

First, the unit of work. Everything on Replicate is a prediction. You create a run. It moves through starting, then processing, then it succeeds, fails, or gets canceled. A run can last up to 30 minutes. That shape fits image and video jobs, where one request in and one artifact out is the whole task. A robot policy is the opposite shape. The loop wants an answer every 20 to 100 milliseconds, many times per second, for as long as the task runs. There is no endpoint that answers with a stream of action chunks. The prediction lifecycle, laid out in the how-it-works docs, is a job API. A control loop is not a job.

Second, the boot time. The same docs say models get turned off when nobody uses them. A cold model can take several minutes to fetch and load before the first answer returns. Cold boots hit less-used models more often. A robot policy on a shared market would be the least-used model on the page. Their fix for cold boots is a paid one. You reserve a deployment with a minimum instance count, and you pay for every second it stays on.

Third, the meter. Public models bill by the second of active processing, per the billing docs, at hardware rates quoted on the pricing page read October 26 2026:

Per-second metering is a fair deal for a 20-second render. It is a poor deal for a loop that needs dozens of tiny calls per second. Each call queues behind other customers on shared hardware. Any cold spell adds minutes. And if you bring your own model as a private one, the meter widens. Private models on dedicated hardware bill for setup time, idle time, and active time. A loop that runs eight hours a day pays for the other sixteen.

None of this says the platform is bad at its real job. It is the best-known place to run a media model for a few cents. The point is narrower. The economics and the API shape were built for batch-shaped work. A control loop is not batch-shaped work. That is why the category is empty. An empty category here says nothing about the models themselves.

What to do instead: four working paths

The models you were searching for are real, open, and runnable this week. Here are the paths, in the order a builder usually tries them.

Path one: run the open weights yourself. Pi 0.5 ships in the openpi repository. One command starts a policy server on port 8000. A client streams action chunks to the robot over a websocket. The repository's own memory table asks for more than 8 GB of GPU memory for inference. It names an RTX 4090 as the example card. A LoRA tune wants more than 22.5 GB. A full tune wants more than 70 GB. OpenVLA, a 7-billion-parameter policy built on Llama 2, ships in its own repository with a light script for serving it over a REST API. SmolVLA trains and runs through the LeRobot stack on cards a hobbyist already owns. For what hardware to buy or rent per family, the hobbyist compute budget guide carries the memory classes and prices. The Jetson versus cloud guide works the onboard tradeoff. If the term VLA itself is new, start with the VLA explainer.

Path two: rent GPUs by the hour for training bursts. Robot policies learn in short training runs, not in steady use. Renting beats owning at small scale. Interruptible marketplace hours for a 12 GB card run a few cents. One rented hour trains the light policy classes several times over. The budget guide linked above carries the worked math and the checkpoint rules that make interruptible hours safe. Keep live control on hardware next to the robot. Rent for the bursts.

Path three: the new hosted robot specialists. As of October 26 2026, two small companies will run robot policies for you over an API. Both are young. Abstract Inc lists six robot models on one API, with a public rate card on its pricing page: a community pi 0.5 tune at 0.85 per hour at a 30 Hz loop, MolmoAct2 at 0.99, their Picasso model at 2.02, Rho at 1.00, and two more at 4.05. All six read today. Robotic Inference lists 24 models from 13 families behind one API, including SmolVLA, GR00T N1.7, OpenVLA, and pi 0.5, with self-measured benchmark scores. Access runs by invitation in batches while it onboards. No public rates yet. Physical Intelligence, the lab behind pi 0.5, opened its own API beta on October 13 2026. It is a waitlist, with batched onboarding, fine-tuning on their machines, and no published prices. Three doors. None of them is Replicate. All of them dated.

Path four: the priced reasoning tier. Google's Gemini Robotics ER 2 is the one big-platform robot API with public rates, read on Google's pricing page October 26 2026. Input runs 1.00 per million tokens and output 5.00 through December 31 2026. Both double on January 1 2027. It is the reasoning layer of a robot stack, spatial judgment and task planning, not the action loop itself. The Gemini rate card guide walks it line by line.

And this platform? Compare options on the pricing page and open an account at signup. The models index is the live catalog. What is on it today is on it today. Reading it beats trusting any page's date, including this one. The wider census of which big hosts carry robot models at all lives in the big hosts guide.

Will Replicate add robotics models?

This section is speculation. It is labeled as such, and dated October 26 2026.

The structure implies it is possible but not signaled. Anyone can push a private container of an open VLA to their own account today. The docs describe exactly that flow. A public, listed robot-policy category is a different thing. It would need per-model economics that survive per-second billing and cold boots. Nothing in the collections index, the blog, or the changelog hints at one being built. The category has been empty for as long as people have searched this way, more than a year.

The honest framing for a buyer: do not block a project on it. If a robotics collection ever appears, the self-check below will show it in under a minute. The working paths above will still be the better first move for anything with a real control loop.

Check it yourself in one minute

Every claim above decays. Re-date it yourself with three checks. They need no account and take less than a minute.

Open replicate.com/collections. Scan the category list for robotics, or anything action-shaped. Then open replicate.com/collections/robotics directly. Watch which error arrives. Then open the pricing page and search it for the word robot. Three live pages, one minute, and you know whether this page has gone stale. That habit, reading the live catalog instead of trusting a search result, is worth more than any single answer here.

Frequently asked questions

Is there a robotics category on Replicate?

No, as of October 26 2026. The collections index lists 39 categories. None is robotics or VLA. The robotics, robot, and vla category URLs return errors. Control URLs for real categories load fine. The pricing page scans clean for robot terms. One research record from October 25 found the same across eleven large hosts. This is a category-wide gap, not one platform's quirk.

Can I push my own VLA model to Replicate and run it there?

Yes, in a technical sense. The platform's own tool, Cog, packages a model into a container you push to your account. Google's AI Overview for this exact query teaches that path. Know the meter first. A private model on dedicated hardware bills for setup, idle, and active time. The API is a job-shaped endpoint, not a control-loop endpoint. A cold model can take minutes to load. For training bursts it can work. For a loop that needs answers many times per second, the math and the latency both cut against it.

Where can I actually run a VLA model over an API today?

Three doors as of October 26 2026. Abstract Inc serves six robot models with a public per-hour rate card. Robotic Inference lists 24 models behind one API, invite-gated in batches, no public rates yet. Physical Intelligence runs its own beta behind a waitlist, fine-tuning included, prices unpublished. Google serves the reasoning tier, not the action loop, at published token rates. None of these is a big general marketplace. That is the honest state of the market this month.

Is VLA-REPLICA the same thing as Replicate?

No. VLA-REPLICA is a test rig from the University of Texas at Dallas, published May 2026. A light box, an arm, cameras, and common parts. Any lab can rebuild it and grade policies the same way. The word replica refers to copying the test setup. It dominates the results for two of the three queries this page covers. That is why those results look useless. If you landed here from a benchmark result, the evaluation cost guide is the neighboring read.

Does anyone price robot inference per call?

The convention is still forming, which is useful to know on its own. Google prices per token. Abstract prices per hour of loop time. The beta programs publish nothing. The per-call and per-robot-hour shapes, and what each one hides, are worked through in the pricing units guide. To compare any of them against running the weights yourself, start at the pricing page.

Last verified: 2026-10-26