Decentralized Physical AI

Nepher Robotics runs tournaments that produce real, deployable robotics AI — trained policies, complete Isaac Lab External Projects, and standardized OpenUSD environments. Those tournaments are validated through Decentralized Physical AI: a transparent, stake-weighted evaluation network where independent validators score every submission in standardized Isaac Lab environments.

The infrastructure runs on Bittensor Subnet 49, but the product is not the blockchain — it is the AI assets the tournaments create.

What Tournaments Produce

Each completed tournament ships three concrete assets the robotics community can clone, reproduce, and deploy:

Trained policy

Every submission centers on a trained control policy — the learned weights that map observations to robot actions. Winning packages include:

• best_policy/best_policy.pt — the required checkpoint validators load during evaluation
• Training metadata — params/agent.yaml and params/env.yaml snapshots so you know exactly which hyperparameters and environment config produced the weights
• Export-ready formats — top packages may also ship JIT or ONNX exports for faster sim playback or downstream hardware deployment

These are not demo weights. They are policies that survived decentralized evaluation against hidden benchmark scenes — ready to load in Isaac Lab with play.py or integrate into your own deployment stack.

Isaac Lab External Project

Each tournament publishes a full Isaac Lab External Project — the standard NVIDIA pattern for custom robot learning tasks. One repo contains everything needed to retrain, evaluate, and run inference:

Component	What it includes
Task module	source/<module>/ — robot definitions, MDP rewards, observation/action spaces, and registered Gym environments
Training code	scripts/rsl_rl/train.py — RSL-RL training entry point with tournament-specific env IDs (e.g. Nepher-Spot-WaypointNav-Envhub-v0)
Inference code	scripts/rsl_rl/play.py — load best_policy.pt and run the policy in simulation with the same env config validators use
Project scaffolding	pyproject.toml, setup.py, and scripts/list_envs.py for install verification

Reference example: task-spot-waypointnav — a hierarchical Spot navigation task with trainable high-level nav policy and frozen low-level locomotion. When a new tournament opens, its task repo follows this same External Project layout so miners and validators share one reproducible codebase.

Environments (public and benchmark)

Tournaments rely on EnvHub OpenUSD environment bundles so every participant trains and evaluates against the same geometry. Two tiers keep competition fair while preventing overfitting:

Tier	Purpose	Availability
Public environments	Training and self-evaluation — diverse scenes miners use to build generalizable policies (e.g. waypoint-sample-v1)	Download anytime via npcli envhub download
Benchmark environments	Official evaluation — hidden scenes validators activate only during the evaluation period (e.g. waypoint-benchmark-v1)	Validators only; miners never see them before submission

Public scenes let you iterate locally with the EnvHub-integrated task variant (Nepher-*-Envhub-v0). Benchmark scenes are the ground truth: if your policy ranks #1, it generalized to environments it was never trained on. All bundles are OpenUSD-compliant and import directly into Isaac Sim 5.1+ / Isaac Lab 2.3+.

After review, podium packages — policy, External Project, and linked environment configs — are released according to each tournament's winner access mode (public download or paid unlock), and selected winners are listed on SimStore, accelerating community learning and enterprise PoCs.

Focus on outcomes

If you are a robotics engineer, start with the Tournament Miner Guide and Tournaments overview. You do not need deep Web3 knowledge to compete — the CLI handles wallet signing and submission, and the podium earns Subnet 49 emissions automatically through validator weights.

Why Decentralized Validation?

Robotics benchmarks fail when evaluation is opaque. A centralized leaderboard can hide dataset leakage, inconsistent environments, or unreproducible scoring.

Decentralized Physical AI solves this by design:

• Multiple independent validators run the same public eval-nav code on each submission in headless Isaac Sim, inside isolated GPU sandbox containers
• Stake-weighted score aggregation — aggregated_score = Σ(score_i × stake_i) / Σ(stake_i) — makes collusion expensive
• On-chain eligibility snapshots lock who can be scored before evaluation begins (Subnet 49 registration + agent submission required)
• Hidden benchmarks — EnvHub evaluation scenes like waypoint-benchmark-v1 are only activated for validators during the evaluation period
• Open-source evaluators let miners self-test with identical configs before submitting

The result: when a policy ranks #1, the community can clone the repo, download the EnvHub scenes, and reproduce the score.

How It Works (High Level)

 Engineers                    Validators                  Output
 ─────────                    ──────────                  ──────

 Train policy  ──submit──►   Download agent ZIP
 in Isaac Lab                 Install task module
       │                      npcli envhub download …
       │                      Run eval-nav (headless)
       ▼                      POST score → api.nepher.ai/tournament
 Tournament task repo                  │
 (USD + Isaac Lab)            Stake-weighted aggregation
                                       │
                                       ▼
                              Admin review → podium (1st–3rd)
                              Podium → Subnet 49 emissions
                              Assets → downloads, highlights, SimStore

Tournament phases (contest → submit → evaluation → review → reward) are documented in the Incentive Mechanism guide.

Bittensor Subnet 49 — Infrastructure, Not the Story

Bittensor provides the decentralized coordination layer:

• Subnet 49 hosts Nepher Robotics validators who evaluate agents and set on-chain weights every 30 minutes
• Wallet registration ties submissions to cryptographic identity (btcli subnet register --netuid 49) — every upload is signed with the miner's hotkey
• Stake-weighted averaging aggregates validator scores; ties break on earliest submission time
• Emission schedule: during a reward period the approved podium receives the reward weight (default 70% / 20% / 10%); each running tournament's preliminary leader receives a fixed 1%; anything unassigned burns on UID 0

Miners interact with the subnet through nepher-cli (npcli tournament submit); validators run the Docker validator from nepher-subnet. Most day-to-day work — training, EnvHub downloads, self-evaluation — happens entirely in Isaac Lab.

For reward distribution details see the Incentive Mechanism; for submission fees and deposits see Platform Balance & Fees.

Comparison: Hackathon vs Tournament

Both live on the Nepher platform but serve different goals:

	Hackathon	Tournament
Purpose	Innovation — explore new task genres	Optimization — best policy for a fixed benchmark
Scope	Broad, open-ended challenges	Specific robot + environment + metrics
Validation	Expert + stake-weighted community voting	Decentralized eval-nav evaluation
Rewards	USD prizes via Stripe Connect	Subnet 49 emissions to the podium
Output	Ideas, PoCs, new genres	Production policies, code, and assets
Portal	nepher.ai/hackathon	nepher.ai/tournament

Get Involved

Role	Guide
Compete (train & submit policies)	Miner Guide
Validate (run evaluations, set weights)	Validator Guide
Browse assets	SimStore · EnvHub
CLI	Nepher CLI

External Resources

• Bittensor documentation: docs.bittensor.com
• Bittensor GitHub: github.com/opentensor/bittensor
• Nepher subnet repo: github.com/nepher-ai/nepher-subnet
• Public evaluation: github.com/nepher-ai/eval-nav
• Tournament API: https://api.nepher.ai/tournament