Robot fine-tuning service: what it is and who is in it
A robot fine-tuning service adapts a robot's policy to your own demonstrations and serves the result somewhere you can call. Two different jobs hide behind the phrase. One is mechanical tuning of control parameters on the machine. The other is machine learning tuning of a policy that steers it. Most of the pages that rank for this phrase describe the first job and sell no robot fine-tuning service. This page names the category. It maps who was in it on October 26, 2026. It sets out the three units the category is priced in.
Last verified: 2026-10-26
Two different jobs hide behind the phrase
Google's own question panel makes the split for you. Asked about robot fine tuning, the panel asks this: "Is your goal physical parameter calibration or AI/vision policy fine-tuning?" On a neighbouring query it asks "Are you fine-tuning a machine learning model (VLA / Diffusion Policy) or mechanical/joint calibration?" The search engine has to ask. The two jobs share a phrase and nothing else.
The first job is control work. You set torque limits, speed profiles, sensor offsets and loop gains. It needs no dataset, no model and no graphics hardware. It happens on your machine, usually with an integrator's hands on it. The glossary entry that ranks near the top for this phrase describes exactly that job. It lists the drawbacks every integrator will quote back to you. The work is complex. It is slow. The result depends heavily on the quality of the data you collect.
The second job is model work. You gather demonstrations, then train a policy on them. You get a checkpoint back, and you need somewhere to run it. That is the job this page covers, because it is the job that can be bought as a service.
The split matters commercially, because the two jobs come from different suppliers and carry different units. Control work is bought from a robotics integrator and quoted per day or per robot. Model work is bought from a machine learning shop and quoted per training run or per delivered policy. Mixing them is the most common way a buyer ends up with a quote that compares to nothing.
One line separates them. If the seller cannot tell you which checkpoint came out of your demonstrations, you are buying control work.
What a fine-tuning service actually sells
Four products are sold under this one phrase. Knowing which one you are buying settles most of the argument before it starts.
| Product | What you get | How sellers sell it |
|---|---|---|
| A tuned checkpoint | The weights, and you host them | Bundled with the endpoint |
| A served endpoint | An address that turns frames and an instruction into motor commands | Bundled with the checkpoint |
| A dataset | The demonstrations, which are the raw material for somebody else's training run | Sold on its own |
| An evaluation report | A counted run of your task, with the tuned policy and the base policy both running it | Rarely bundled |
Most sellers bundle the first two. Physical Intelligence's API beta does exactly that. Its announcement, published October 13, 2026, gives the flow. Upload data, fine-tune on compute the lab manages, then serve the tuned models through a hosted gateway (announcement). The checkpoint sits inside the endpoint in that case. The evaluation report is usually missing, and that is the part worth asking for.
Ask one question before you send a dataset or a budget. Which of the four am I buying? A seller who answers in one sentence has thought about the category. A seller who answers with a paragraph about machine learning has not.
Who was in the category in October 2026
The field is small, new and split by leg. This section is dated October 26, 2026. Read every line as a snapshot rather than a standing fact.
The model lab came first. Physical Intelligence opened an API beta on October 13, 2026. The beta makes its latest models available for training and inference (announcement). The workflow has three steps. Upload data, fine-tune on compute the lab manages, then serve the tuned model through a hosted gateway. Teams enter in batches through an application. No public rate card exists for the beta.
Two specialist hosts then appeared. Abstract sells hosted robot models and takes fine-tuning work as a separate engagement. Its rate card, read on October 26, 2026, prices each model per million tokens in and per million actions out. It adds an hour estimate for a 30 Hz control loop. Those rates cover serving. Robotic Inference sells one API across pi 0.5, GR00T, X-VLA, SmolVLA and OpenVLA. Its own description of the workflow has four steps. You pick the model that fits the task, then fine-tune it on your data. You test it in simulation, then run it on the robot. Its landing page states no rate at all.
The data side is its own business. Firsthand collects custom egocentric video for embodied AI and robotics. It bills per validated hour of footage, and its landing page, read on October 26, 2026, reports 10,585 validated hours delivered. That is the input leg of the category, priced per hour of recording rather than per model.
Then there is the do it yourself path. Amazon published a walkthrough of fine-tuning OpenVLA on SageMaker AI with LoRA in June 2026 (AWS Physical AI blog). That is the same job a service sells, done on rented GPUs by your own engineers. Two details from the writeup shape the bill. LoRA updates fewer than 1% of the parameters and freezes the rest. The resulting policy cut action prediction error by 44% against the base model on a held-out set.
The method research explains why a service sells judgement as well as compute. The OpenVLA-OFT paper puts the position plainly. Vision language action models "struggle with novel robot setups and require fine-tuning to achieve good performance, yet how to most effectively fine-tune them is unclear given many possible strategies" (paper, 2025). The recipe it proposes combines action chunking, parallel decoding and a continuous action representation. It reached an average success rate of 97.1% across four suites of the LIBERO benchmark (project page). Choosing between recipes is the judgement, and that judgement is a large part of what a service sells.
Much of what ranks for this phrase belongs to language models. Fine-tuning as a service for text models is a mature product. It has its own price lists and its own cloud vendors. None of it touches a robot. One row that ranks for the head query makes the point on its own. The "Fine-Tuning as a Service, TaaS for AI Models" post on rafay.co returned "Page Not Found" at one read on October 26, 2026 and served its full text at a second read the same day. The post sells tuning as a service for language models. It never touches a robot.
The last signal is the quietest and the most useful. On the same day, Google returned no suggestions at all for seven commercial stems. That covers "robot fine tuning service", "robot fine tuning service pricing", "hosted vla fine tuning" and "robot policy fine tuning service". It also covers "vla fine tuning service", "robot model fine tuning" and "fine tuning service for robots". It returned suggestions only for the plain "robot fine tuning", and the top suggestion there was an academic paper title. An empty suggestion box means the commercial vocabulary for this category has not formed yet. That is why every page that ranks answers a slightly different question than the one you asked.
The three units the category is priced in
There is no single price for a fine-tuning service. The category bills in three different units, and mixing them up is how budgets go wrong.
| Unit | What it covers | What moves it |
|---|---|---|
| The training run | One fine-tune of one policy on one dataset | Model size, demonstration count, step count, and full against adapter training |
| The delivered fine-tune | A checkpoint, and ideally a counted evaluation on your own task | The size of the model and how much of it the run touches |
| The robot-hour | The serving of the policy that comes out | Calls per second, the rate per call, and the hours the robot runs |
The training run is the first unit, and it is the one that is easiest to misjudge. The bill tracks four things. They are the size of the model, the number of demonstrations, and the number of training steps. The fourth is whether the run updates every parameter or a small adapter. The adapter choice is the largest lever on that bill. Full training and an adapter differ by more than a hundredfold in parameters trained. It is also the least costly decision to get right. The AWS writeup shows the adapter path reaching a good policy while touching under one percent of the parameters.
The delivered fine-tune is the second. What you buy is a checkpoint plus, ideally, proof that it beats what you started from. That proof is a counted evaluation on your own task, not a demo video. A vendor who quotes the training run but not the evaluation is quoting you half the product.
The robot-hour of the resulting policy is the third, and it is the recurring number. It is not a training number at all. It is inference spend: calls per second, times the rate per call, times the hours the robot runs. On this platform the action lane meters by the call. A robot-hour is your loop rate times the per-call rate times your hours.
The shape worth remembering is simple. The training leg is paid once, while the serving leg is paid every hour the robot runs. A quote that covers only the first number leaves out the one that keeps arriving.
No number is typed on this page by design. This site measures the train and evaluate loop and the per robot-hour cost separately. Both measured pages are listed with this page. The train and evaluate arithmetic sits on the fine-tune and eval product math guide. The serving side sits on the measured robot-hour cost guide. Live rates for the lanes this platform serves render on the pricing page.
How to tell a service from a tutorial
Five questions. Ask all five before you send a dataset or a budget.
Does the seller name the artifact? A checkpoint, an endpoint, a dataset or an evaluation report. If the answer is a method name, you are reading a tutorial.
Does the seller show a before and after on your task? A tune with no counted evaluation on the task you care about is an opinion.
Does the seller quote a unit you can plan with? Per training run, per delivered policy or per robot-hour. A quote with no unit cannot be compared with another quote and cannot be budgeted.
Where do the demonstrations live, and who else may use them? Ask directly. The dataset is often worth more than the tune. A vendor who will not say who holds it has not thought about your second tune.
Is there a way off the service? Can you export the checkpoint and run it on hardware you own? A vendor who will not let you leave is selling a subscription, not a fine-tune.
Where this platform sits in the chain
The chain has four steps: collect demonstrations, fine-tune, evaluate, serve. The first three are one purchase, and the fourth is a different one, often from a different company.
This platform sits on the fourth step. It serves vision language action models over an action route and meters by the call. The robot inference API category head maps that serving tier in full. As of October 26, 2026, the eight rows in the platform's model data for that route are not on the public list. That set holds the pi 0.5 family with its DROID and LIBERO tunes. It also holds SmolVLA, OpenVLA-OFT, ACT, X-VLA and the MolmoAct2 arm tune. The models index is the live catalog, and the action lane page carries the request shape.
This platform does not publish a robot fine-tuning service. If you want the serving step without owning the hardware, that is the part you can buy here. If you want the fine-tune itself, the guide listed with this page maps the hosted options. The open tooling does it yourself when you would rather run the training. The alternatives to the hosted fine-tune guide covers that side.
Getting started
Four steps, and the first one is a decision rather than a click.
- Open a console account. Start at the signup page. One account covers every lane the platform lists, and the wallet, the keys and the spend history live there.
- Read the rates once, in one place. The pricing page renders every listed row's rate from the platform's own model data. The pricing guide explains the units and how the prepaid wallet behaves.
- Decide which leg of the chain you are buying first. The dataset, the fine-tune or the serving. They have different suppliers and different units, and the serving is the only one that recurs.
- Read the guides listed with this page. They cover the measured product math for the train and evaluate loop, the alternatives to the hosted fine-tune, the measured per robot-hour cost, the local versus hosted break-even, the evaluation benchmark, the policy server and the pricing unit glossary.
The category is young enough that its vocabulary is still forming. Start from the unit and the artifact, not from the phrase.