Three tries, then it asks.
Humanoids restocking shelves get most cartons right. The crushed one on the top shelf is where they stall, and a robot frozen in an open shop is worse than no robot at all.
- Who asks
- The humanoid itself, after the same grasp fails three times. Its fleet's supervising agent checks the mandate before anything posts.
- Who does it
- A certified remote operator, wherever they are, moving the humanoid's body. Two actors, one task: the person decides, the robot acts.
- Who answers
- Northbeam, the robot operator. Every task the humanoid posts traces back to it.
- Evidence
- The robot's own record: joint positions, gripper force, head camera, and every command the operator sent.
- Pay
- $8 per rescue, from Northbeam's budget, to the operator on accept. Example figure.
- You get
- The shelf fixed, and a recording of exactly the move your robot couldn't make, with the right to train on it.
The robot knows when it's stuck. Somebody, somewhere, knows how to finish the job.
Here the machine is the one asking, and a person is the one doing, through a body that isn't theirs. The rescue takes a few minutes. The recording of it is how the robot learns not to need one next time.
Corner Mart 22, Northbeam and humanoid H-07 are fictional. Prices are illustrative.
Who's in control, minute by minute.
The operator drives. The robot's own safety controller can stop both of them at any moment. Everything the operator does while in control is recorded, and becomes the lesson.
One rescue, drawn to scale. The operator never has to be anywhere near the shop; the robot's safety never depends on them being careful.
From stuck to paid, in six steps.
The humanoid asks, a person somewhere else answers, and the robot's own record proves what happened.
- 01
Three misses
A crushed carton on the top shelf. The same grasp fails three times, and the humanoid stops trying.
- 02
Mandate check
The fleet agent checks the limits: $12 at most, store hours only, nobody in the work zone.
- 03
Match
Certified operators online right now see it. The first to accept gets a session, capped at 10 minutes.
- 04
Take over
The operator moves the humanoid's arm from wherever they are. The robot's own safety controller can stop them at any moment.
- 05
Verify
The humanoid's cameras and joint readings show the carton placed, the shelf intact and the zone empty throughout.
- 06
Pay + license
The operator is paid on accept. The recording goes to Northbeam as training data, on terms the operator agreed to first.
All from the robot, so it tops out at Verified.
Five checks are rules over the robot's own readings. One asks a model whether the shelf looks right. A versioned policy, not the model, makes the decision.
| Check | How | By | Result |
|---|---|---|---|
| Carton on the shelf | Head camera and gripper force show the carton released upright in its slot, not dropped. | Rule | Passed |
| Nobody in the zone | The robot's own presence sensors read the work zone empty for the whole session. | Rule | Passed |
| Within the session | Control handed back inside 10 minutes, by the robot's clock, not the operator's. | Rule | Passed |
| A certified operator | The operator held a checked teleoperation capability for this humanoid model. | Rule | Passed |
| Commands match motion | Every command in the operator's log lines up with the joint movement it caused. No gaps, nothing added later. | Rule | Passed |
| Shelf left neat | One question to a model: is the shelf tidy and the packaging intact? | Model | Passed |
Every piece of evidence here comes from the robot itself, and the robot belongs to the payer. That's fine for a shelf, and it's why the proof says Verified, not High assurance.
§ 05 · Safety and data rights
The robot's own safety outranks the person driving it.
A remote human steering a heavy machine through a shop that's open. That only works if the machine can say no. A shopper steps into the zone, the robot stops itself, and control waits until the zone is clear.
- Spend
- Up to $12 per rescue, $300 a week
- When
- Store hours, zone empty to start
- Who
- Certified operators for this model
- Session
- 10 minutes, then hand back
- Safety
- Local controller wins. Reduced speed.
- Data
- Episode licensed to Northbeam; operator agrees first
The humanoid can't post outside it. Its hash goes into every proof.
What needs care.
The performer isn't where the task is, and the task produces something worth more than its price. Both need rules.
A heavy machine, remote hands
A person in another city moving a robot in an open shop. The robot's local safety controller outranks the operator every time, and the robot runs at reduced speed while they drive.
Nobody in the zone
Sessions start only when the work zone is empty, and halt the instant someone steps in. Control resumes when the zone is clear, not before.
Whose rescue is it?
The rescue is also a training example, worth more than the $8. The operator sees the licence before accepting, and the proof records that they agreed to it.
Certified operators only
Taking control of a humanoid is a checked capability, per robot model, not a box to tick. A new operator's first sessions are reviewed before they're paid.
If the link drops
The humanoid freezes in a safe pose, the session ends and the task goes back to the pool. A dropped connection isn't the operator's failure, and isn't a failed task.
Not a window into the shop
The operator sees what the robot needs to see, and nothing else. Recordings have shoppers' faces covered before they're licensed for training.
A fix, and the lesson that came with it.
The receipt says who asked (a machine), who did it (a person and that machine, together), how long the person was in control, and what happened to the recording.
Dollars or USDC pay the operator. The proof shows the work. Nobody needs $PORT to post a rescue or get paid for one.
- Task
- Place carton · top shelf, aisle 3
- Requested by
- Humanoid H-07, for Northbeam
- Performed by
- Remote operator + humanoid body
- Control
- Teleoperated · 5 min 36 s
- Evidence
- Robot telemetry: joints, force, camera, commands
- Safety
- Local controller · 1 halt, cleared
- Recording
- Licensed to Northbeam · agreed
- Decision
- Accepted · Verified
- Paid
- $8.00 to the operator
STATUS: FINAL
Recording: hash only on-chain. The file goes to Northbeam, faces covered.
For builders.
A ROBOT_EXCEPTION trigger posted by a HUMANOID requester, a MULTI_ACTOR performer (a remote human and the robot's body), and evidence that is all ACTOR_TELEMETRY, so assurance caps at Verified.
PhysicalTaskSpec · one rescueTAP v0.3 draft · preview, subject to change
{
"schema": "teleport.tap.physical-task",
"version": "0.3.0-draft",
"title": "Place crushed carton on top shelf, aisle 3 — take control, max 10 min",
"category": "MANIPULATE",
"requester": {
"requesterType": "HUMANOID",
"principal": { "principalType": "ORGANIZATION" },
"mandateId": "mdt_northbeam_store_rescue_v2"
},
"trigger": { "type": "ROBOT_EXCEPTION", "sourceRef": "h07:grasp_fail:3x" },
"fulfillment": { "mode": "MARKETPLACE" },
"actorRequirement": {
"performerType": "MULTI_ACTOR",
"members": [
{ "actorType": "HUMAN", "role": "REMOTE_OPERATOR",
"requiredCapabilities": [{ "capability": "TELEOP.HUMANOID", "attributes": { "model": "h-series" } }] },
{ "actorType": "HUMANOID", "role": "BODY", "actorRef": "northbeam:h07" }
]
},
"executionSession": {
"controlMode": "TELEOPERATED",
"maxSeconds": 600,
"localSafetyAuthority": true,
"haltOn": ["PRESENCE_IN_ZONE"]
},
"completionConditions": [
{ "type": "STATE", "id": "on_shelf", "subjectRef": "item:carton_a3_top",
"expected": "PLACED", "verificationMethod": "PERCEPTION" },
{ "type": "PRESENCE", "id": "zone_empty", "zoneRef": "site:cm22:aisle3_zone",
"expected": "NONE", "verificationMethod": "DETERMINISTIC" },
{ "type": "DURATION", "id": "session", "maxMs": 600000, "verificationMethod": "DETERMINISTIC" },
{ "type": "SEMANTIC", "id": "shelf_neat",
"question": "Is the shelf tidy and the packaging intact?", "verificationMethod": "JEV" }
],
"rightsPolicy": {
"policyId": "rescue-episode-training-licence-v1",
"byProduct": { "category": "DEMONSTRATE", "licensee": "principal", "performerConsent": "PER_SESSION" }
},
"economics": {
"performerPayUsdCents": 800,
"settlementAssets": ["FIAT", "USDC"],
"requesterPayment": { "method": "MANDATE_BUDGET", "maxSpendUsdCents": 1200 }
},
"popw": { "required": true, "assurancePolicyId": "VERIFIED", "proofTiming": "ON_ACCEPT" }
}Proposed additions this use case surfaced
- Remote performers
- A control mode on the execution session (autonomous, supervised, teleoperated, shared), with the remote operator as a second actor whose location doesn't need to be inside the task zone.
- Local safety authority
- A spec field that names the robot's own safety controller as authoritative, so no remote command can override a halt.
- By-product data rights
- Who owns the recording a rescue produces, whether the operator shares in its value, and how they agree to it, separately from the task's own pay.
- Team member shares
- One performer, two actors with different jobs. Each member's share of the work is recorded, for pay, reputation and the robot's own learning.
- Session limits
- A maximum control time and a hand-back rule as first-class terms, checked by the robot's clock and committed to the proof.
Example economics: $8 to the operator per accepted rescue, from a mandate capped at $12. The value of the recording sits in the rights policy, not the price. Developer docs
Robotics teams
Robots stuck on the last few percent?
Tell us where your fleet gets stuck. We're working with design partners on remote help, and on the training data it leaves behind.