OpenRoboto

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

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.