Mythbusting Physical AI: Thoughts from a recent panel

When Amazon announced their one-millionth robot deployment this summer, it seemed like a watershed moment. But Amazon is the exception, not the rule. Most logistics operations don't look anything like Amazon's hyper-structured environments.

At Staer, we build intelligence for mobile robot fleets operating in the reality most enterprises face: messy operations, ad hoc processes, and under-instrumented facilities. Recently, I joined Søren Halskov Nissen from Yaak and Sam Baker from Planet A for a panel at The Drop, where we unpacked hard truths about mobile robotics. Several persistent myths aren't just wrong—they're actively holding our industry back.

Myth 1: Synthetic Data Is Enough

There's a seductive narrative: skip the messy work of collecting real-world data and train models in simulations. But without ground truth from the real world, synthetic data just amplifies your blind spots.

Real-world data from deployed machines beats everything if you can get it at scale. Many vision language models train on "demo mode" operations—perfect conditions, ideal lighting. Deploy a robot trained on that into an actual warehouse under throughput pressure, and it'll move like an intern on their first shift.

Myth 2: Open Source Isn't Defensible

There's still a reflex to vertically integrate and build everything in-house. But this wastes effort. I've watched engineers spend months reinventing tools that already exist in robust, open-source form. There are components that are just easier and better built in the open.

When you're getting your SDK installed on millions of robots, transparency becomes your greatest asset. Integrators and customers trust what they can see and understand.

The real moat isn't proprietary code. At Staer, we focus on three compounding assets:

Data: How we collect it, and the pipelines we build.

Evaluation: Testing models against out-of-distribution scenarios that prove genuine generalization.

Deployment: Making robots hit throughput targets in messy, unstructured environments.

Myth 3: Task-Specific Autonomy Equals Generalized Intelligence

This might be the most damaging myth. Until your robots perform boringly well in environments they've never seen before, you don't have a solution that scales.

We're moving toward smaller, specialized models working side by side. You load what you need—manipulation, navigation, mapping—rather than one bloated network. It's more efficient, flexible, and easier to iterate.

The path is clear: collect expert demonstrations in real conditions, train on curated benchmarks, then test on completely unseen environments.

What Actually Matters

A few more myths worth killing:

Compute isn't the bottleneck. A modern VR headset runs all the algorithms you need for a mobile robot. Robots fail because models are inefficient and brittle, not because they lack processing power.

Hardware churn doesn't matter. Form factors are stabilizing. The real upgrade path is continuous model improvements and safe over-the-air updates.

Looking Forward

By late 2026, non-technical people will deploy robots using natural language interfaces. You'll instruct machines in plain language, like children. This transformation is already happening in software. Robotics will follow.

This arrives at the right moment for Europe. We face urgent productivity challenges, shrinking labor pools, and automated global competitors. For many industrial players, there's no alternative but to automate this decade.

At Staer, we're solving one of the critical deployment friction points: enabling robots to operate in constantly changing, unstructured environments. The next million robots won't all live in Amazon fulfillment centers. They'll operate across Europe's factories, farms, and logistics networks—in the messy, unpredictable real world where they're needed most.