An open competition on Bittensor for continuously improving robotics models.
SN 80. Rank 9 of 32 by emission
OpenRoboto
Emission3.0%of the block reward · rank 9 of 32
Participants5 miners · 9 validators
Registration0.0005 TAO
Verified2026-09-09links checked and scores merged
The programtraining · github.com@726e42a · read 2026-09-19
The work
A fine-tuned robot control model, uploaded whole to Hugging Face. Miners start from the season's base model (LingBot-VLA 2.0 for simulation since 2026-09-01), train it by any recipe, publish the complete merged checkpoint, pay a fee in TAO, and announce the Hugging Face commit on chain. Every submitted model is public and free to download.
own docs · raw.githubusercontent.com
publish their fine-tunes as complete model checkpoints on Hugging Face
How it is scored
A benchmark worker runs each checkpoint through LIBERO manipulation tasks in the MuJoCo simulator, seeded from the commitment's block hash plus a drand beacon. The score is the average success rate. Ranking is king-of-the-hill: a challenger takes the crown only if its score exceeds the champion's by more than 0.01; a tie loses and the challenger is left off the board. The evaluation code is in a separate repository.
own docs · raw.githubusercontent.com
the challenge succeeds only if challenger_score > champion_score + champion_margin (default 0.01)
Sources disagree The overview at the pinned commit says the round scores six LIBERO suites for a pi0.5 base; the live competitions API on 2026-09-19 names the simulation benchmark libero_pro_custom_1 with a LingBot-VLA 2.0 base and a baseline of 0.285833 over 5 trials.
How the pay splits
Three pools. Simulation takes 15 percent of the miner share and pays only its top three at 70, 20 and 10 percent of that. Each of two real-robot tracks on an xArm 6 takes 42.5 percent, settled once a season: 95 percent to the single champion vested over 120 days, 5 percent shared by entries clearing the qualification bar. Live weights on 2026-09-19 read 0.425, 0.425, 0.105, 0.03, 0.015.
own docs · raw.githubusercontent.com
relative weights 70 / 20 / 10 within the simulation track's 15% emission allocation
Who judges
Nine validators were active on the 2026-09-09 chain snapshot, of 64 permitted. Weights are committed then revealed, at most once per 100 blocks. Validators do not run the benchmark: validator.py reads a settled weight vector from the operator's backend API with a read-only key, normalizes it and calls set_weights. The evaluation itself runs on the operator's benchmark workers.
chain · hyperparams-2026-09-19.json
commit_reveal_weights_enabled True, weights_rate_limit 100, max_validators 64; chain-2026-09-09.json active_validators 9
How often it turns
Emission settles every 360 blocks, about 72 minutes. The validator's default is to set weights every 720 minutes. The simulation board turns per submission: a checkpoint is scored when it reaches the front of the queue and the champion holds until beaten. Real-robot tracks run as seasons of roughly a month, with fixed submission and evaluation windows. A new registration is immune for 5,000 blocks, about 16 hours.
chain · hyperparams-2026-09-19.json
tempo 360, immunity_period 5000; validator.example.yaml weight_interval_min: 720
Who buys
Nobody is named. The front page describes a competition whose winning model becomes the next base, and every checkpoint is public on Hugging Face. The competitions API lists fees paid by miners, 0.1 TAO burned for simulation and 1.5 TAO transferred for a real track, and no price for anyone else. No customer, product or usage figure was found on the site, in the API or in the repository.
our reading · openroboto.ai
Fine-tune the base model, submit your weights on-chain, and the best model becomes the next base.
What has been gamed
The overview records a red-team study: the four base LIBERO suites could be saturated for about $50 of GPU time by memorizing public demonstrations, so two swap-perturbation suites were added. Other named attacks and answers: queue flooding, met by the burn fee and one commitment per hotkey per round; copying the champion, met by the tie-loses margin; resubmitting another miner's weights, met by shard fingerprinting; replayed burns, a 10-block window.
own docs · raw.githubusercontent.com
the four base suites alone could be saturated for roughly $50 of GPU time by memorizing public demonstrations
2 of 7 fields rest on the chain or the code. A manifest is the site's reading of the subnet's program, not a review; each field links its source. All programs · What a field means · Correct a field
What it says
The subnet in its own words: the line the index took from its front door, and what the team has said on the record.
What it shows
Explained from zero by the index, from the subnet's own materials, for someone who has never heard of Bittensor.
SN80 OpenRoboto runs a sheepdog trial for robot brains, where every handler trains a pup from the same litter, the course is unseen until the day, and the champion keeps its place until a dog clearly beats it.
The commodity, explained from zero
The output is a model checkpoint: a saved file of a robot's control brain, the kind downloaded to make a robot arm pick things up when told to in plain English. These are called vision-language-action models because they take in camera images and a text instruction and put out arm movements, as a sheepdog takes in a field and a whistle and puts out a run. The subnet starts from a public base model, the shared litter, and pays people to train a better dog from it. Miners "Fine-tune the official base model" and then "Publish a complete merged checkpoint to Hugging Face and submit on chain."
The strongest robot models today, the front page says, "are built inside a few private labs." Whether anyone will pay for an open one is unstated. No customer is named; the only hint at a buyer is a front-page line about hosting a "policy tournament based on real factory need." The champion's file is public on Hugging Face, so a buyer would pay nothing to download it.
Why it is on Bittensor at all
What the network adds is an open, checkable scoreboard and a purse. Anyone can enter, the course is drawn from randomness no handler can predict, and every score comes with an audit record. Here the comparison breaks: a champion sheepdog goes home with one handler, while here the champion is a file anyone can copy, and the prize is the subnet's token, vesting over 120 days.
How the work gets done
Miners fine-tune the season's base model (LingBot-VLA 2.0 this season), upload the full checkpoint, pay a fee in TAO (0.1 burned for the simulation track, 1.5 transferred for a real-robot track), and announce the submission on chain. Validators run each checkpoint through LIBERO manipulation tasks in a physics simulator, a trial field made of software, with test seeds derived from the block hash plus drand, a public randomness beacon, so they are "unpredictable before submission, frozen after it, and reproducible by anyone." A challenger wins only if its score exceeds the champion's by more than 0.01; a tie fails. Emission, the subnet's share of newly minted TAO, is split 15 percent to the simulation board and 42.5 percent to each of two real-robot tracks on an xArm 6, the trials run on a real field, where 95 percent of the track's pool goes to the single champion.
How you would know it works
The public competitions API at https://api.openroboto.ai/api/v1/competitions lists three active competitions, their fees, and a simulation baseline score of 0.286. The leaderboard at https://www.openroboto.ai/#/benchmark renders only with JavaScript; the index's fetch on 2026-09-09 saw a champion at 0.841.
What is missing
There is no product, price, or API for buyers, and the on-chain contact riccardo@openroboto.ai appears nowhere on the site or in the repositories. The chain's github_repo README calls its own submission flow deprecated and points to a second repository. The site is an empty page without JavaScript. No Discord is set on chain; support is "Open an issue on GitHub."
Go deeper
- Protocol and incentive mechanism document: https://raw.githubusercontent.com/openroboto-ai/openroboto-subnet/main/docs/SUBNET_OVERVIEW.md
- Current miner guide, dated 2026-09-02: https://raw.githubusercontent.com/openroboto-ai/openroboto-cli/main/docs/MINER_LINGBOT.md
- Live competitions and fees API: https://api.openroboto.ai/api/v1/competitions
Sources for this explainer
- https://www.openroboto.ai/
- https://www.openroboto.ai/#/docs
- https://www.openroboto.ai/#/benchmark
- https://api.openroboto.ai/api/v1/competitions
- https://raw.githubusercontent.com/openroboto-ai/openroboto-subnet/main/README.md
- https://raw.githubusercontent.com/openroboto-ai/openroboto-subnet/main/docs/SUBNET_OVERVIE
- https://raw.githubusercontent.com/openroboto-ai/openroboto-cli/main/README.md
- https://raw.githubusercontent.com/openroboto-ai/openroboto-cli/main/docs/MINER_LINGBOT.md
Metaphor: a sheepdog trial for robot brains. Every claim is drawn from the evidence set or the subnet's own materials; "(inferred)" marks a conclusion rather than a quote. Corrections.
How legible it is
Four audiences, four questions each, rated on what a first-time reader can find in five minutes. The scale measures findability, not quality.
On-chain identity 5/7- subnet_name
- OpenRoboto
- github_repo
- github.com
- subnet_contact
- riccardo@openroboto.ai
- subnet_url
- openroboto.ai
- discord
- not set
- description
- An open competition on Bittensor for continuously improving robotics models
- additional
- not set
Stakers and validators●●●●● 3.3
Should I allocate here? · rank 12 of 32 for this audience
Q1What it is●●●●● 3
An open competition on Bittensor for continuously improving robotics models.
matches the chain description word for word, output is open model checkpoints on Hugging Face, no unit, no price
Q2Who it is for●●●●● 3
3 ranked submissions refreshed Sep 10
champion 0.841 dated Sep 09, 48 submissions on the front page, the arena telemetry panel on the same page reads 0 submissions, no customers or revenue
Q3How it resists gaming or fails●●●●● 4
named Protocol and Incentive Mechanism document one click from the README with an Anti-gaming summary, seeds from block hash plus drand, plagiarism check, a red team note on memorization
Q4Identity and documentation●●●●● 3
identity 5 of 7, discord and additional not set, github and url resolve, contact address not on the site, a For Validators page and VALIDATOR.md exist, no staker or tokenomics page
Miners●●●●● 4.0
Can I compete, and what wins? · rank 3 of 32 for this audience
Q1What it is●●●●● 4
Fine-tune the official base model
then "Publish a complete merged checkpoint to Hugging Face and submit on chain" with the required checkpoint layout and a CLI walkthrough
Q2Who it is for●●●●● 4
Linux, an NVIDIA GPU (24 GB VRAM minimum) and a recent driver
plus Python 3.11 and Docker, fee 0.1 TAO burn (sim) or 1.5 TAO transfer (real) from the competitions API, registration cost not stated
Q3How it resists gaming or fails●●●●● 4
challenge rules in prose, a challenger must beat the champion by more than 0.01, top 3 weighted 70 20 10 within a 15 percent simulation share, real tracks pay 95 percent of a pool to the champion
Q4Identity and documentation●●●●● 4
guide updated 2026-09-02 with a what works today table, CLI on PyPI 1.1.0 or newer, cli repo pushed 2026-09-09, the chain github_repo README says its own rt.py flow is deprecated, no GitHub releases
Buyers and enterprises●●●●● 2.0
Can I use this today? · rank 20 of 32 for this audience
Q1What it is●●●●● 3
Pull the base model, fine-tune it, submit the weights on-chain.
outputs are open checkpoints and an Open Data Pool, no product, API or app for buyers
Q2Who it is for●●●●● 1
Host specific policy tournament based on real factory need
is the only mention of a customer, no customer type statement and no pricing anywhere
Q3How it resists gaming or fails●●●●● 2
Open an issue on GitHub, we read everything
is the support path, no status page, SLA or response expectation, no Discord on chain
Q4Identity and documentation●●●●● 2
a read-only public API for competitions, submissions and weights is listed in section 10, the chain contact riccardo@openroboto.ai appears nowhere on the site or in the repos
Newcomers●●●●● 3.8
What is this and why does it matter? · rank 9 of 32 for this audience
Q1What it is●●●●● 4
Pull the base model, fine-tune it, submit the weights on-chain. A model only replaces the base if it scores better.
on the front page under the chain description, fine-tune and on-chain unexplained
Q2Who it is for●●●●● 4
Most of the strongest robot models today are built inside a few private labs. We build them in the open.
openness versus private labs stated on the front page, no cost or coverage comparison
Q3How it resists gaming or fails●●●●● 4
live leaderboard with scores against a 0.286 base, submissions dated Sep 05 to Sep 09 and per row audit links, the telemetry panel above it reads zero, page renders only with JavaScript
Q4Identity and documentation●●●●● 3
subnet_name matches the site, url resolves, description is a noun phrase without a verb, no about page, the docs and roadmap sections are the only learn pages
Provenance
How this rating came to be. Verified means the scorer fetched the page and the words are on it; secondhand means concluded from code, absence, or a third party. The link badge is a separate automated check made before publication.
scored by agent a 2026-09-09, rubric v1.0merged 2026-09-09links verified 2026-09-09: 36 live, 0 unreachable, 0 manual
Something wrong? Corrections of fact are applied as they arrive during the window; rating disputes are batched at its close. How to file one · GitHub issue · email.