Roboarm VLA

Workload class: text-directed vision–language–action (VLA) manipulation — bi-manual and kitchen-grade scenarios are common examples; the subnet is not limited to food service.

On testnet, operators issue task prompts; miners compete on trajectories and policies from VLA / control stacks; validators score candidates on safety, task match, and efficiency. Outputs feed Proof-of-Physical-Work (PoPW) records on the Konnex testnet.

Testnet

The public dashboard URL for this workload will be announced with the testnet release.

Task shape (illustrative)

The live UI and chain define the exact fields. Example natural-language tasks:

"Pick the apple from the pan and place it on the counter."
"Dice tomatoes and place them into the pan."

A JSON sketch:

{
  "jobId": "sha3(signed-packet)",
  "prompt": "dice tomatoes and place into pan",
  "arena": "KitchenSim-v1",
  "deadline": 120,
  "rewardTestKNX": "…",
  "kpi": {"time_s": 90, "success": true, "spills": 0}
}

Use the fields and token type shown in the testnet product (e.g. testKNX).

Miner output

Policy or VLA bundle (e.g. WASM or integration endpoint) with deterministic seeds where the subnet requires them
Declared KPIs and stake per network rules

Validation flow (conceptual)

Deterministic replay in the subnet’s simulator or harness when applicable.
KPI extraction and safety checks.
ScoreRoot emission; rewards or slashing per onchain parameters.

Example prompt

"Pick the red bell pepper, slice into strips, sauté for 2 minutes, then place on the plate."

Good night