LeRobot hosting: what exists today and what does not

As of October 26 2026, no provider sells hosted LeRobot as a product category. This page is the receipt for that sentence. It names every host checked. It says what each one sells instead, and how each check was run. Everything is dated today. The short map: the training side of LeRobot is already commercially hosted. The serving side is unsold under this name. That side means a provider who keeps your trained policy online and rents you the endpoint. If you searched the phrase, you met the collision. The results explain where LeRobot keeps its files. The project readme itself says datasets are "hosted on the Hugging Face Hub". That answers a storage question. Nobody on the results page answers the buyer's question. This page does.

Last verified: 2026-10-26

What LeRobot is

LeRobot is Hugging Face's open-source library for teaching robots by example (LeRobot). You guide a robot arm through a task a few dozen times. The robot learns the motion from those demonstrations. The library spans cheap 3D-printed arms up to humanoids. It gives you the whole loop in one toolbox. It records demonstrations. It stores them in a standard dataset format that streams from the Hub. It trains policies with one command. It runs the trained policy on the real robot with another (LeRobot docs). The project's paper calls it an open-source library for end-to-end robot learning (paper, 2026). The span runs from low-level control through data tooling to the learning methods. The serving half of the pattern has its own deep definition in the policy server guide on this site.

What hosting LeRobot would actually mean

Hosting LeRobot would mean two products, not one. The split is the key to the whole category.

The first product is hosted training. Your demonstrations go up. Graphics time gets rented. The trained checkpoint comes back. This product exists, first-party. The LeRobot hardware guide documents it. Add one flag to the ordinary training command. The job lands on Hugging Face Jobs, on managed machines, billed by the minute (hardware guide). The LeLab app wraps the same path in a point-and-click interface. It walks you from an unboxed arm to a trained policy (LeLab docs).

The second product is hosted serving. A trained policy is a file. The file wants a graphics card and an always-on address. The library splits that job in two by design. A policy server loads the checkpoint on a machine with a graphics processor. Your robot computer keeps the cameras, the motors, and the recording. That exchange repeats all day. The robot sends its camera frames, its joint positions, and the instruction. The server sends back a chunk of future actions, plus the scaling numbers the policy was trained with. The robot keeps moving through the chunk it holds while the next chunk computes. Today's engine ships a command named lerobot-policy-server as its serving entry point. The deployment command, lerobot-rollout, is the single CLI for running trained policies on real robots (policy deployment docs). Hosting this half means someone else keeps that server alive, metered, and pointed at your robot. That is the product nobody sells under this name.

The memory classes set the size of the machine, and they are published (hardware guide). The small ACT class wants about 2 to 6 GB of video memory. SmolVLA wants about 10 to 16 GB. The pi0 and pi0.5 class wants about 24 to 40 GB. A hosted offer would pick one card size per policy family. The buyer would never think about it again.

The census: who was checked, and what they sell instead

Every row below was checked on October 26 2026. The method is named per row. An absence claim is only as good as its check.

Two rows deserve plain words, because they are the closest the market comes.

Hugging Face was checked in three pieces. The Hub hosts the artifacts: datasets, checkpoints, demo apps. Jobs hosts the training, with the same lerobot-train command, by the minute. The serving check came back empty. The model list API was queried for any LeRobot-format model with a hosted inference provider attached. The answer was zero, on a model whose own tag is robotics. Training is a product there. Serving is not.

The two specialist rows are new in 2026 and real. abstract.inc sells a fixed menu of robot models behind one API. It prints rates per hour at a 30 hertz loop. Its pi 0.5 community tune sits at 0.85 an hour, read October 26 2026 on its own page. roboticinference.com sells one API over 24 models from 13 families. It fine-tunes on your data and tests in simulation. It takes teams by registration or pilot. Neither mentions LeRobot on any page fetched. Neither offers to hold your lerobot-train checkpoint online. They sell their menu. The category this page names is the one thing they do not sell. That would be someone else's machine, holding your checkpoint, under the ecosystem's own name.

What people do today instead

With no category to buy, the community self-hosts at every layer. The costs are known.

At the bottom, buy the card. A used 3060 12 GB sold for about 295 in October 2026. A used 3090 sold for about 1,093. Those are sold prices, not listings, from the data gathered on the hobbyist budget guide on this site. The 3060 trains the light ACT class end to end. The 3090 reaches every open policy family at small batch sizes. The full ladder, with the memory wall per family, is that guide's own subject.

In the middle, rent by the hour for training bursts. In our own October 2026 training runs on one community market, a mid-priority 3060-class hour billed about 0.04. A 3090-class hour billed about 0.041. Advertised rates on that market start at 0.015. Training suits rented machines because checkpoints resume. Live robot control does not suit them. The placement guide on this site draws that line by timescale (Jetson versus cloud).

At the top, managed training. Hugging Face Jobs bills 1.00 an hour for its small 24 GB card. The 80 GB A100 bills 2.50 an hour. Both were read October 26 2026 on the official pricing page (Jobs pricing). The same command you run locally submits there. The SO-101 guide on this site walks one arm through the whole path, with real commands for the middle shapes (SO-101 cloud brain).

What nobody rents you is the always-on endpoint. That gap is the category.

The security bar any host would have to clear

A hosted offer puts a policy server on a network strangers can reach. That fact has a dated lesson attached.

The United States vulnerability database records CVE-2026-25874 against the LeRobot async policy server and robot client, through version 0.5.1 (NVD). The channels carried no sign-in and no encryption. The components unpacked data from the connection without checking it. A network-reachable attacker could run code of their choosing on the machine holding the graphics processor. Independent security writeups followed through 2026. The exposure proof is that people were running these servers where strangers could reach them at all.

The project's answer reshaped the serving path. The async module is being removed from the main branch. Its replacement merged on October 12 2026 and ships the lerobot-policy-server command. The new engine's own documentation warns that binding an interface limits exposure but does not authenticate peers. It points to certificate-based transport for anything outside a trusted lab network. For you, the rule is short. Keep your server on a network you control. For a host, the bar is higher still. A real offer authenticates every peer and says so in writing.

Why the category is empty, and why it will not stay so

Three reasons, all dated.

First, the demand is young. Google's autocomplete returns nothing for "lerobot hosting". It returns nothing for "hosted lerobot". It returns nothing for "lerobot cloud". All three were checked today through the suggestion service. The organic results are library docs, one paper, one explainer, and forum threads. Forum threads about running LeRobot in production sit on page one. The question exists. No seller has planted a flag for it to find.

Second, the phrase collides with a different meaning. The search engine's own summary reads the query as a storage question. It explains that LeRobot datasets and models live on the Hub, then asks whether you meant to upload a recording. A buyer typing the phrase gets a filing answer. A category has to break through that before it can even be searched for.

Third, the economics closed only recently. The measured record on this site put a full fine-tune and evaluation loop at 2.08 to produce in October 2026. The written charge band for such work runs 15 to 40. The product math guide on this site shows that arithmetic in full. When the underlying work costs a single-digit percent of that band's midpoint, a packaged hosted version becomes worth building. The entry demand is the skill loop: record demonstrations, tune lightly, serve the result. The pricing unit is also unset. One specialist platform prices per hour at a fixed loop rate. One frontier vendor prices by the token. No per-call or per-robot-hour convention exists yet. Whoever serves the loop first names the category and sets its unit.

What a real hosted LeRobot offer would include

If you are evaluating any offer that markets itself this way, or waiting for one, this is the checklist.

1. Your checkpoint, not their menu. The service loads the policy you trained, from your fine-tune, on your robot's data. A fixed catalog of publisher models is a different product. Useful, but not this one.
2. The normalization numbers travel with every answer. A policy tuned with one set of scaling numbers and served with another moves the arm wrongly, without erroring. The answer must carry the numbers it applied.
3. A stated chunk contract. Chunk size, refill point, and blending decide whether motion is smooth. The offer publishes them. You tune them per task.
4. Authenticated transport. The 2026 record for this server class is unauthenticated channels and unchecked loading, ending in remote code execution. The replacement engine documents certificate-based transport. A hosted offer has no excuse to do less.
5. A metering unit that fits loops. Per call or per robot hour, stated in the open. A robot at ten calls per second makes about 36 thousand calls an hour. Any per-call rate published anywhere becomes an hourly bill you can now compute.
6. A live catalog for state truth. What serves today, what does not, checkable by machine, dated. The section below applies that law to this platform itself.

Where this site stands

This platform publishes a hosted action route for robot policies. One call goes in: camera frames, an instruction, and the robot state as numbers. One chunk of actions comes back, with its normalization numbers. The route contract is public on the action route reference. The VLA lane page carries the models and the call contract in full. As of October 26 2026, the action route's eight rows are not on the public list. No rate for them renders anywhere. The models index is the live catalog. Rates for listed rows render on the pricing page. How metering and the prepaid wallet work is on the pricing guide.

The category this page names is the one this platform is built around. Open robot policies, served behind one metered endpoint, priced in units a robot loop can plan on. If you want the endpoint side to exist, open a console account at the signup page. Create a key. Point the client you already have at the published contract. The guides listed with this page carry the rest. One defines the policy server. One runs the full host census. One walks the self-host ladder, one works the per-hour cost math, and one covers the LeRobot camera path. The category fills in from there.