Overview

Start here

Attune, built by Robotensor, is a Bittensor subnet for adaptive robot intelligence: models that learn new work from what they are shown, with their weights frozen. Vector, its first competition, is open now.

2 min read

Adaptive robot intelligence

Robots meet work that keeps changing: new objects, new layouts, new products. The usual answer is to collect data and retrain. Attune pays for the alternative, adaptive robot models: models that are shown the work at deployment and do it, with their weights frozen. Whatever such a model works out lives in its context, a demonstration or a video, and nothing in its weights changes.

Attune measures that capability in the open. Miners train models, validators settle them in head-to-head duels on scenes no one could have seen when the model was committed, and every result is published for anyone to check.

The competitions

Competition	The context a model is shown	What it is asked to do	Status
Vector	One robot demonstration of a task, with its actions	Do the same task in a new scene, where the objects have moved and the other arm may have to act	Open
Horizon	A human or robot video, and the robot's own history	Predict the future physical state and the robot's actions	Coming soon

Vector in one paragraph

Every entrant trains the weights of one fixed in-context network, vector_v1.1, so a duel compares training (data, curricula and methods), never architecture or code. A challenger plays the reigning model on an identical list of units dealt from a chain block finalized after it committed, and takes the crown by beating the king's score by +3.0 pts. No miner code ever runs, and every duel is published unit by unit, with its clips.

Where to go next

VectorIntroduction

What Vector measures, the model in one picture, and where the crown stands.

Vector · The guideThe benchmark

The sixteen tasks, what one unit is, and how the scenes will widen.

Vector · The rulesWhat you submit

The weights, the repository and the on-chain commitment, step by step.

Vector · OperatingRunning a validator

One GPU host, three environments, and the commands that run them.

Every protocol number these pages quote is read from the competition's contract, spec.json, the same file the validator runs on, so no page can disagree with the rule it describes.