robotics training data, labelled by a network.

Robots learn hands-on tasks from video, but only once every step is labelled. Thor Machines runs a Bittensor subnet where miners break first-person footage into timed, captioned steps and validators check the work. We are building the platform and the subnet code behind it.

15–20%

hidden test jobs, known answers

Pull

miners fetch jobs when ready

open footage → open dataset

What we built

the system.

Task order book: open and completed annotation jobs, with sample videos
Bittensor wallet sign-in: miners identified by their on-chain key and whitelisted to apply
Annotation tool: upload video, extract frames, mark when each step starts and ends
Revenue page tracking the subnet's wallets on-chain
Miner and validator code: a pull-based job queue, scored against hidden test videos
Validator checks with vision-language models, scored on catching deliberately broken labels

Stack

BittensorNext.jsSupabaseVercelVision-language models

Outcome

quality, not method.

Miners can label by model, by hand or both: the network rewards accurate steps, timings and captions. V1 is in development on open-source footage, building toward an open dataset that shows the network can do this work well.

robotics training data, labelled by a network.

the system.

your project,next.

your project,
next.