Teleoperation network · sapienscale.com

Humans as an API.

SapienScale is a teleoperation network. Your robot requests a certified human operator over an API; we attach one in under a second — from Manila, Delhi, Da Nang, or Nairobi — and carry their hands to your machine at the speed of glass.

→ Run a human for the whole session, or attach one only when autonomy taps out — billed by the second, either way.

< 1 sOperator attach
≈ 148 msClosed-loop latency
11,000 kmProven glass reach
4 regionsOperator pools
[01]

Continuous or on-demand

Run a human the whole session, or let robots run their own policy and attach one only for the edge cases autonomy can't close — the exception handler, on demand, then detached. Same network, same latency, either way.

[02]

Physics is fixed. Throughput isn't.

We don't beat the speed of light and won't pretend to. The engineering problem is everything around it: encode time, head-of-line blocking, congestion collapse, reconnects. That's what we solve.

[03]

Labor is where it is

Skilled operators exist at global scale; robots that need them mostly don't share a timezone. A transport that holds up over 11,000 km of fiber makes location irrelevant.

§1

The network

Operator pools connect to the nearest relay; robots connect to theirs. The relay fabric is on the physical path, not a detour — media and control ride the same submarine fiber your packets would take anyway. Operator side →

FIG 1 — operator POPs, relay fabric, robot sites · edge labels are one-way time in glass
§2

Transport: independent streams

Every sensor is its own stream; control is unreliable datagrams. A lost video packet can't stall steering — there is no shared queue to stall. This is the whole trick, and it's why the loop stays tight on lossy trans-oceanic paths. Deep dive →

FIG 2 — media path left→right, control path right→left · inset: loss behavior, ordered pipe vs independent streams
§3

Autonomy ↔ human handoff

Two modes on one substrate. In continuous mode a human drives the full session; in on-demand mode autonomy runs and a human attaches only below the confidence threshold τ, handing back above it. Both command against a digital twin the robot keeps synced to the console — operators act on the twin, the robot reconciles. How handoff works →

FIG 3 — handoff state machine · every transition is logged and becomes training data
§4

The honest latency budget

Manila to San José is ~11,200 km — 55 ms one-way in fiber, and no protocol changes that. What we control is everything else: the loop closes in ~148 ms, inside the 200–250 ms a human needs to react to what they see. Full physics →

FIG 4 — one-way budgets to scale · dashed marker: full video→reaction→control loop
§5

Trust & safety

You're handing a remote human real-time control of a physical machine. The design answers the four questions that come next before you ask them.

CERT & SCOPE

Operators are credentialed per skill

Every operator is certified for a specific skill and machine class. A session only matches a pool whose certification covers your robot — an L3 forklift job never lands on an untrained pair of hands.

FAIL-SAFE

Link loss stops the machine, not the logic

If the operator link degrades or drops, the robot falls back to a policy hold and slows to safe — the digital twin keeps the last validated state. No operator, no motion.

ACCOUNTABLE

Every transition is logged and attributable

Attach, command, and handback are signed and recorded. You get a full audit trail per session — the same log that becomes training data is also your incident record.

DATA & REGION

Your facility feed stays encrypted and scoped

Media is encrypted end to end. Session video is scoped to the matched operator for the session envelope and isn't retained outside your region without consent.

§6

Request a human

One call. Specify the skill, the machine, and the session envelope. The pool is matched by certification and measured RTT to your robot's relay.

POST /v1/sessionscurl
curl https://api.sapienscale.com/v1/sessions \
  -H "Authorization: Bearer $SAPIENSCALE_KEY" \
  -d '{
    "robot_id": "amr-042",
    "skill": "forklift.pallet_recovery",
    "max_rtt_ms": 180,          // hard ceiling, measured to relay
    "handoff": "twin_synced",   // operator attaches against digital twin
    "max_duration_s": 600
  }'

// → 201 · operator attached in 640 ms · rtt 112 ms · cert forklift/L3
§7

Join the network.

Two sides, one API. Put your machines on the network, or get paid to run them.

FACILITY OWNER

Put your robots on the network

Tell us the machines and the skills. We'll scope a pilot on your fleet and get you an API key.

Request access →
OPERATOR

Get paid to operate, from anywhere

We're onboarding certified teleoperators in Manila, Delhi, Da Nang & Nairobi. Bring your skills; we bring the machines.

Apply to operate →

Prefer email? hello@sapienscale.com