Loading
Loading
Neuron registration trends
Kinitro is a Bittensor subnet that incentivizes the development of generalist robotics policies capable of performing well across diverse simulated environments (MetaWorld manipulation and Genesis humanoid locomotion/manipulation). Miners train policies, deploy them to Basilica serverless infrastructure, and are evaluated using epsilon-Pareto dominance scoring that rewards true generalization across all environments rather than specialization.
Kinitro uses a split service architecture: (1) Miners deploy policy servers to Basilica (serverless GPU) and commit endpoints on-chain. (2) The scheduler reads on-chain commitments to discover miners. (3) Task generators create evaluation tasks stored in PostgreSQL. (4) Executors fetch tasks, run simulations (MuJoCo/Genesis in Docker), and call miner endpoints for actions. (5) Results are submitted back, and scheduler computes epsilon-Pareto weights. (6) Validators poll the weights endpoint and submit to chain. This architecture separates heavy computation (backend operators) from lightweight validation (community validators).
The scoring mechanism: (1) Each miner is evaluated on all environments, computing success rates. (2) For each subset of environments, the miner that Pareto-dominates others on that subset wins points equal to the subset size. Epsilon tolerance prevents noise from determining winners. (3) Points from all subsets are aggregated. (4) First-commit advantage gives earlier miners priority in ties - later miners must exceed the leader's score by a threshold gap (2-10% depending on evaluation samples) to win. (5) Final points are converted to weights via softmax with configurable temperature. This design ensures specialists only win small single-environment subsets while generalists win larger, more valuable subsets.
The sections above were auto-generated from the GitHub repository and may not reflect the latest changes. View repository
Decentralized adversarial robustness network
This information was auto-extracted from the GitHub repository and may be incomplete or out of date. Always refer to the subnet's GitHub repository for the latest instructions.
Holder & total-staked metrics as of 1h ago (hourly snapshot)
GPU required for training robotics policies. For inference, can deploy CPU-only to Basilica for testing but GPU recommended for production. Training typically requires high-end GPU (A100, H100, or similar).
Validators are lightweight - they only fetch weights from the backend and submit to chain. No GPU required.
Initialize a new policy from the miner template
uv run kinitro miner init ./my-policy
Test your policy server locally before deployment
cd my-policy && uvicorn server:app --host 0.0.0.0 --port 8001
Upload your trained model weights to HuggingFace
huggingface-cli login && huggingface-cli upload your-username/kinitro-policy .
Deploy your policy to Basilica serverless infrastructure (required for mainnet)
export BASILICA_API_TOKEN='your-token' && uv run kinitro miner push --repo your-username/kinitro-policy --revision YOUR_HF_SHA --gpu-count 1 --min-vram 16
Register your policy endpoint on-chain so validators can discover and evaluate you
uv run kinitro miner commit --repo your-username/kinitro-policy --revision YOUR_HF_SHA --deployment-id YOUR_BASILICA_ID --netuid YOUR_NETUID --network finney --wallet-name your-wallet --hotkey-name your-hotkey
Install the kinitro package and dependencies
git clone https://github.com/threetau/kinitro.git && cd kinitro && uv sync
Start the validator (polls backend for weights and submits to chain)
uv run kinitro validate --backend-url BACKEND_URL --netuid YOUR_SUBNET_ID --network finney --wallet-name your-validator-wallet --hotkey-name your-hotkey