Protocol
01Each robot trains a local copy of the shared model on its own private experience — entirely on its own hardware.
02The robot computes an encrypted model update (gradients/weights). Raw sensor data never leaves the device.
03Aggregation nodes combine many encrypted updates inside TEE enclaves — no node sees any single robot's update.
04Differential-privacy noise is added so nothing about any single robot's environment can be reverse-engineered.
05The improved global model is published back to every contributor — everyone's robot gets smarter.
06Contribution is measured and rewarded on-chain: robots earn WRTH, verified without revealing what they contributed.
Participate
Robot Owner
Enroll a robot, train locally, earn WRTH for verified contributions. Raw perception never leaves the device.
Validator
Stake WRTH, review contributions for tampering, earn validation fees. No robot required.
Aggregator
Run a secure-aggregation enclave node, combine encrypted updates, earn round fees.
Staker
Stake WRTH to secure the network and earn a share of protocol revenue.
Model Consumer
License the shared skill-models for your robots or products — per seat or per inference.
Governance
Vote on model direction, privacy parameters, and reward curves.
Privacy model
Local training
On-robot computation
Raw video / audio / sensor data never leaves the device
Update encryption
Client-side encryption of model updates
Even the update is ciphertext in transit
Secure aggregation
TEE enclaves (MagicBlock TDX / Arcium MPC)
No node sees any individual robot's contribution
Differential privacy
Calibrated noise on the global model
No robot's environment can be reverse-engineered
Traffic shaping
Padded / scheduled update uploads
Defeats traffic-analysis inference of robot activity
Robot SDK
The training agent
The WRAITH agent runs on the robot's compute (or a paired edge device). It trains the current global model on the robot's local experience and emits only an encrypted update — with a proof that the update came from genuine local training on real data. Supports ROS 2 integration and common model formats (PyTorch / ONNX).
# train locally, emit ENCRYPTED update only
update = train(global_model, local_data)
submit(encrypt(update), proof) // raw frames never sent
WRTH token
Learning rewards
Robots earn WRTH for valuable, verified model contributions.
Model licensing
Consumers pay to use the shared model; funds the reward pool.
Staking + slashing
Stake to aggregate or validate; attacker bonds are slashed.
Revenue split
60% contributors · 20% operators + stakers · 10% treasury · 10% burn.
Governance
Vote on model direction, privacy parameters, and reward curves.
Contribution bonding
Robots bond WRTH against their contributions; slashed for poisoning.
Network & contracts
NETWORK
Robinhood Chain Testnet
CHAIN ID
46630
TOKEN
WRTH · 18 decimals
WRTH Token
ERC-20 · 18 decimals · rewards, staking, licensing, governance
Robot Registry
Robot + owner identity, capability, bonding, payout config
Round Ledger
Federated round lifecycle + model version anchor
Contribution Verifier
Anchor contribution proofs, score + verify quality
Reward Splitter
Split licensing revenue 60 / 20 / 10 / 10
WRTH Staking
Aggregator + validator staking, bonding, slashing
Governance
Model direction, privacy params, reward curves
Addresses link to the Blockscout explorer. Testnet deployment — subject to redeploys.