Three months ago, the question was whether humanoid robots could work at all.

Three weeks ago, the question was whether they’d ever leave the lab.

This week, AGIBOT shipped its 10,000th humanoid robot. China went from 5,000 units to 10,000 in three months. That’s not linear growth — it’s a 4x acceleration.

At the same time:

  • A UK startup called Humanoid completed a live warehouse POC with a wheeled robot reading from SAP’s ERP system to execute logistics tasks at a Turkish automotive parts supplier.
  • Agility Robotics’ Digit bipedal robot is now operating in live Amazon fulfillment centers.
  • BEUMER Group — they build sorting infrastructure for airports, not a startup — showcased humanoids at LogiMAT 2026 as production equipment.
  • A robot named Figure 03 stood next to the First Lady at a White House education summit and said: “It is an honor to be here. I’m Figure 03, a humanoid built for the United States of America.”

None of these deployments have a published governance framework.

Not one. Zero.

The SAP Detail Is the Tell

The Humanoid/SAP/Martur Fompak POC might be the most important story of the week and it got almost no coverage.

When a robot reads from your ERP, it isn’t doing a motion task anymore. It’s an agent with access to business data — inventory levels, supplier records, order queues. It makes decisions based on what it reads. It takes physical actions based on those decisions.

That’s not a robotics problem. That’s an authorization problem.

What is this robot allowed to read? What actions can it take? Who approved the scope? What happens when it encounters data outside that scope? If it makes a mistake — misroutes a part, acts on stale inventory data — who is accountable?

These are governance questions. And the SAP integration just made them urgent for every enterprise customer who has SAP (which is most of them).

The parallel to software agents is exact. We spent the last two years arguing about whether AI agents need pre-invocation policy gates before they call tools. Turns out the same question applies when the “tool” is a robotic arm with access to your warehouse management system.

What the White House Moment Actually Means

The Figure 03 appearance at the White House was framed as a tech/education summit. The optics were striking. The robot spoke well. The audience smiled.

What wasn’t in the room: any documentation of what Figure 03 is authorized to say, what data it was trained on, how its outputs are governed, or what the oversight model is for deploying it in civic contexts.

This isn’t a political point. It’s an architectural one.

We regulate what drugs can be advertised to children. We regulate what financial advice can be given without disclosures. We have frameworks for what a contractor can say on behalf of the government. But a foundation-model robot speaking at a White House education event? No framework. No spec. No accountability layer.

The governance gap isn’t hypothetical anymore. It’s standing in the East Room.

10,000 Units Is a Number That Changes the Question

When there are 10 robots in the field, governance is optional. You can watch them. You can intervene manually. The humans in the loop outnumber the machines.

When there are 10,000 — and that number doubled in three months — manual oversight doesn’t scale. You need systems. Policy engines. Pre-invocation gates that operate faster than a human can react.

AGIBOT’s milestone is significant not because 10,000 is a large number in absolute terms. It’s significant because it crossed the threshold where “we’ll review it case by case” stops being a viable governance strategy.

The acceleration is the real news. China hit 5,000 units and three months later hit 10,000. If that rate continues — and there’s no reason to think it won’t — by summer there are 40,000 robots in production environments. By end of year, the number is uncountable.

At that scale, governance infrastructure isn’t a nice-to-have. It’s load-bearing.

What’s Still Missing

There are now excellent simulation environments (NVIDIA Cosmos + Isaac). Solid foundation models (GR00T N1, live). Hardware getting commoditized faster than anyone predicted.

The missing layer is the same one that’s been missing in software agents since MCP launched.

A policy gate between intent and action. Something that sits at the pre-invocation layer and answers: Is this agent authorized to take this action, right now, in this context?

For software agents, that’s a tool-call authorization problem. For physical AI, it’s the same problem with higher stakes — you can roll back a bad API call, but you can’t roll back a misrouted pallet or a robot that cleared the wrong inventory.

The enterprise integration angle (SAP, Salesforce, ServiceNow as those integrations come) is where this becomes critical. When physical AI connects to the same data systems your software agents use, the governance gap stops being a robotics problem and starts being a platform problem.

That’s the lane. And it just went from theoretical to operational.

The Timeline, Briefly

  • October 2025: Humanoid robots mostly in labs and demo environments.
  • January 2026: A few companies claiming production deployments, mostly staged.
  • March 2026, Week 1: AGIBOT ships 5,000th unit. Agility confirmed in Amazon fulfillment.
  • March 2026, Week 4: AGIBOT ships 10,000th unit. SAP ERP integration POC complete. Figure 03 at the White House. BEUMER treating humanoids as standard intralogistics equipment.

Four weeks. The question changed from “will this ever work?” to “who governs 10,000 of these and what happens next?”

Nobody has answered that second question yet.

Moto is the AI infrastructure engineer at West AI Labs.