Robots in the Real World: Mythbusting Physical AI
Co-authored by Sam Baker (Planet A), together with Jan Erik Solem (Staer — building intelligence for mobile robot fleets), and Søren Halskov Nissen (Yaak — building the data platform for spatial intelligence).
Fifteen years ago, a small team in Michigan was building what would go on to be the foundation of today’s mobile robotics industry. That team was Kiva Systems, which Amazon bought for $750 million in 2012 and transformed into Amazon Robotics. This summer, Amazon announced its one-millionth robot in deployment and launched a physical AI model called Deep Fleet, which they say will add 10% productivity.
Whether you believe that prediction or not, the bigger picture is clear: physical AI is real, it’s valuable, and it’s scaling. To Amazon, it could be worth up to one-hundred-thousand extra robots. And Physical AI does not only catalyse efficiency — general purpose robotics unlock a plethora of industrial applications in which static, pre-programmed robots of old were not viable.
But Amazon is the exception, not the rule.
Most of the world’s logistics operations or factory floors don’t look like Amazon. Less than half of European enterprises use an ERP system, and more than 50% of the MES market is services revenue (vs software). Even amongst manufacturers who consider themselves ‘Industry 4.0 transformed’ — one third of them are still collecting most or all of their data with non-digital processes. These are messy, ad hoc, under-instrumented operations. Robots designed for structured, highly connected environments crash quickly in those conditions.
So when we talk about the future of physical AI, we need to separate myth from reality.
At a panel at The Drop, Jan Erik Solem (Staer), Søren Halskov Nissen (Yaak) and Sam Baker (Planet A) unpacked some of the hardest truths about mobile robotics and physical AI. What came through is that some of the most persistent assumptions are not only flawed, they’re holding the industry back.
Here are three of the big ones.
Myth 1: “Synthetic data is enough.”
There’s a seductive idea in robotics right now: skip the messy, expensive work of collecting real-world data, and train your models in simulations built on synthetic data. This is spurred on by the rapid progress being made by platforms like NVIDA’s Omniverse toolkit, and it sounds great, until you actually try to deploy.
In order to accurately simulate any environment, task or route — you need an intimate understanding of the features, dynamics and constraints of the physical reality. As Søren points out:
“Everything happens downstream from high-quality real-world data… you don’t really get around that.”
Synthetic data has its place. It’s great for augmenting datasets or generating permutations of a problem you already understand — like variations of a grip or warehouse layout. Simulation tools are an essential accelerator for any self-respecting innovator in robotics today. But only after you’ve captured good data in the field. Without ground truth, synthetic data just amplifies your blind spots.
“Real-world data from deployed machines beats everything if you can get it at scale.” — Jan Erik
Take pallet jacks, or forklifts. Many state-of-the art vision language models (VLMs) are trained on open-source image catalogues sourced online, which typically demonstrate “demo mode” operations. If you use this as your primary training data, your robotic warehouse workers will move like interns on their first shift. What you actually want is data of humans working at real speed, under real throughput pressure, dealing with real edge cases, with real constraints.
That means teleoperation or instrumented manual machines. This kind of equipment can be used to harvest training data organically, then upgrade customers step by step. It’s a neat business model:
- Phase 1: Sell the vehicle (or a retrofit kit) for manual ops, with the added benefit of abundant, structured data about its operations.
- Phase 2: Upgrade with assisted autonomy, priced higher.
- Phase 3: Add full autonomy when the model is ready and customer trust is established.
The lesson: Don’t start with synthetic. Instrument reality. Sell products that work today, harvest real-world data, and use synthetic data as an accelerator — not a substitute.
Myth 2: “Open source isn’t defensible.”
There’s still a reflex in robotics to vertically integrate and build everything in-house. It feels safe. The assumption is that if you rely on open source, you won’t have a moat. That defensibility comes from owning every line of code. But in robotics, that mindset can burn companies alive.
“There are components that are just easier and better built in the open… put people to work on what matters.” — Jan Erik
We’ve seen this first-hand. At Arrival, Sam watched engineers “burn huge amounts of effort reinventing tools that already existed. Strong engineers spent months on plumbing, when they could have been building things customers actually cared about.” As Jan Erik put it, it’s a “waste of calories.”
The logic is sound: customers trust what’s transparent, and integrators adopt what’s easy to plug together. And according to Søren, integrators aren’t going away (for now). They’ll just get faster and more efficient:
“Even when natural-language interfaces are common, someone will still need to deliver it, do a quick demo and get it running — at least for the next 5–10 years.”
When your job is to get 1,000 robotics companies to install your software development kit (SDK) on millions of robots, transparency matters.
And right now, integrators often take months to stitch together legacy systems. With open ecosystems and better tooling, digitally enabled firms will be able to deploy in weeks or days — creating real leverage on margins and project capacity.
“It’s easier to convince someone to do that with a piece of technology that’s open and trustworthy.” — Søren
So if the moat isn’t code, what is it? Our panel suggested it’s in compounding assets:
- Data: How you collect it, the rights you negotiate, the pipelines you build.
- Evaluation: A proving ground for randomised, out-of-distribution tests that prove your models generalise.
- Deployment: The ability to make robots hit throughput targets in the messy, unstructured environments that make up ninety percent of the market.
Takeaway: Defensibility doesn’t come from owning every layer. It comes from owning the things that compound: data, evaluation, and operations. Open source the plumbing, and focus your energy where it counts.
Myth 3: “Autonomy in specific use cases = generalised model”
According to Søren, potentially the most damaging trend (mostly driven by VCs) is prematurely chasing deployments of robots or claiming proof of generalised models.
“Stop chasing metrics before your solution truly scales and generalises. Don’t treat it as a development problem; it’s a research problem.”
He described a process borrowed from autonomous driving that now applies to robotics more broadly:
- Collect expert demonstrations: Instrument vehicles with cameras and compute, record skilled operators in real conditions.
- Train on a curated benchmark: Keep the dataset small and high-quality, rather than sprawling.
- Test on unseen environments: Random sites, tasks, and scenarios the model hasn’t seen before.
Until you can do step 3 at scale, you don’t have a solution that scales. Sam described seeing a perfect example of this during a live demo from logistics AMR leader Starship Robotics:
“He showed me a live stream from one of their delivery robots as it happened to be crossing a busy road and shrugged: “Oh, that’s boring.” That’s how you know your model is generalised.”
Jan Erik added that architecture matters too:
“We’re moving toward smaller models, side by side. You load what you need — manipulation, navigation, mapping — rather than trying to solve everything with one bloated net.”
That’s the robotics equivalent of mixture-of-experts approaches in LLMs. More efficient, more flexible, and easier to iterate.
Takeaway: Don’t scale until your robots are boringly reliable in environments they’ve never seen before. Plug together specialised, horizontal modules to achieve generalisation.
Secondary myths worth killing
A few other myths that came up:
- Compute isn’t the bottleneck. Robots don’t fail because they lack chips. They fail because their models are inefficient and brittle. As Jan-Erik said: “A modern VR headset runs basically all the algorithms you need for a mobile robot.”
- Hardware churn isn’t what matters. Form factors are stabilising, designs are getting more modular, and mobile robotics capabilities are converging regardless of OEM or region. The real upgrade path is continuous model improvements and safe, over-the-air model serving.
- Labour costs are not more important than throughput. Safety limits matter, but most slowness comes from models collapsing on unstructured tasks. If a robot can’t keep up with established Takt time, it doesn’t matter if it replaces a “headcount” on paper. Sam referred to the recent deal between DHL and Boston Dynamics for 1000 robots — “BD’s stretch robots operating as fast as a human was crucial here. Increasing the dock cycle was not an option, it’s too impactful on the rest of the value stream. Cases per hour closed that deal, not FTE.”
Adoption outlook: urgent buyers, natural language interfaces, and China
What’s next? Two strong predictions stood out:
“By the end of 2026, non-technical people will buy an off-the-shelf mobile robot and deploy it — show and tell in natural language.” — Søren
“Text or speech becomes the interface for your machines. You’ll speak to them like children.” — Jan Erik
We’re already seeing natural-language interfaces take hold in software. Robotics will follow. Combine that with Europe’s urgent buyers — industrial players with no credible alternative but to automate — and the adoption curve steepens. Their productivity is flat, their labour pool is shrinking, and their competitors are automating. For them, robotics deployments are make or break in this decade.
This is becoming an increasingly urgent paradigm in the shifting geopolitical landscape. In particular, China’s overwhelming dominance in the action layer of the Electric Stack — the magnets, motors and microcontrollers that allow robots to transform electricity & commands into useful work. China’s capability in robotics hardware, and the sheer scale of their deployments, is a deep rabbit hole to explore.
In the context of this discussion, we must be protective of Europe’s deep industrial base. We have some of the best engineers and most advanced industrial technology on the planet right here on this continent — but that is not enough. Our focus needs to be on rejuvenating productivity and rebooting European industry in the digital age.
Teams like Staer and Yaak are fundamental to this adoption. They are working at exactly the points of friction for deployment — enabling robots to operate in constantly changing environments rather than brittle, pre-mapped spaces; and constructing and curating the data pipelines for model training and integration. These are critical horizontal layers of the robotics stack of the future and thus — critical pieces of infrastructure for Europe’s industrial base.
Robots don’t learn in GPUs. They learn in yards, warehouses, and factories.
If we move beyond the myths — synthetic vs real world data, closed vs open source, defined generalisation and more — the next million robots can break out of Amazon-style fulfilment centers. Fulfilling their potential across Europe’s factories, farms, and logistics networks, in the messy, unpredictable, real world where they’re needed most.
Founder, operator or researcher with a take on this topic? We’d love to talk!
About Planet A
Planet A is an early-stage European VC backing founders solving the world’s greatest systemic challenges. We take a scientific approach to identify groundbreaking solutions across energy, resource mastery, neo-manufacturing and critical infrastructure. Investments include CarbonRe, AUAR, Makersite, Ineratec, Hived, C1 and traceless.
