01
Software agents hit a wall at the edge of the machine.
They move money, mint, and post — but every action stops where the physical world begins. There is no POST /wave for reality.
Live demo · Bankr → real humanoid
Software stops at the screen.
Your robot won't.
Type one sentence to Bankr. A real Unitree humanoid moves in the room. Nothing unsafe ever reaches the motors — every command passes a bridge that is fail-closed by design.
1sentence → real-world motion
∞agents, one physical world
0unsafe actions, by design
Say it, and it moves
“wave hello on my Unitree”arm waves
“take two steps forward”robot walks
“turn to face the camera”robot turns
“what's your battery?”reads state back
The thesis
The next interface for agents isn't a screen. It's the world.
02
Physical action should be a callable primitive.
As composable as a token transfer. As rentable as compute. Any agent, any wallet, pays for real-world motion the way it pays for an API call today.
03
Safety belongs at the edge — not in the model.
The agent may ask; the bridge decides. A whitelisted, state-gated, fail-closed runtime stands between intent and any actuator that can hurt someone.
04
Whoever owns that bridge owns the on-ramp to embodied AI.
Unitree is building it in the open — one proven loop at a time, starting with a humanoid that waves from a sentence you type to your assistant.
What we're building
One platform. A growing family of embodied skills.
The Unitree bridge is the safety runtime. Each product plugs a new kind of intent, payment, or hardware into the same fail-closed bridge.
Unitree Skill
Type a sentence to Bankr → a real Unitree humanoid moves. The first end-to-end proof that an assistant's language can become safe physical motion. Runs in MuJoCo sim today, on live hardware next.
View the skill ↗Unitree Bridge
The safety runtime every skill sits on: action whitelist, FSM & battery gating, bearer auth, rate limits — fail-closed. The agent asks; the bridge decides.
Action Market (x402)
Wrap any action behind x402 so any agent can pay USDC per action and get physical work back. The robot becomes a metered, permissionless actuator any wallet can rent.
On-chain Triggers
The bridge subscribes to chain events — tip received, mint, gas threshold — and fires an action with no human and no prompt in the loop. The thing a pure software skill structurally can't do.
More bodies
Arms, quadrupeds, mobile bases — every new machine speaks the same whitelisted vocabulary, so safety scales with capability instead of fighting it.
How it works
You type it. The bridge keeps it safe. The robot moves.
You say it
Plain English to Bankr, or an on-chain event.
AI maps it
Your assistant picks exactly one whitelisted action.
Unitree Bridge
Auth, whitelist, state & battery gating. Refuses anything unsafe.
It moves
A real humanoid moves — on rails, on-chain, fail-closed.
Non-negotiable
Fail-closed is the whole product.
A robot that walks from the wrong state falls over. So the bridge — not the AI — holds every safety-critical decision.
Whitelist onlyOnly named actions can ever run. No improvisation reaches hardware.
State gatingMotion refused unless the robot is in a safe, balanced state.
Battery gatingWalking & turning refused below a safe charge.
Explicit confirmDangerous actions require a deliberate second yes.
Rate limitedCommands are throttled to protect the hardware.
No token, no motionMissing auth ⇒ every request rejected. Closed by default.
The endgame isn't "an agent that can wave."
It's a world where physical action is just another callable primitive — as composable as a transfer, as rentable as compute, and only ever safe. We start with one sentence, one wave, one real robot.