The Robotics and Physical AI Wave

Adriana Vitagliano

Along with Lorcan, Michael and our brilliant team, I invest in the best European founders building at pre-seed across software and hardware

Robotics is having a moment

 At least that’s what it feels like if you look at the VC funding numbers, with ~$27 billion deployed into robotics in 2025 - more than double the 2024 amount (per Pitchbook).

At firstminute, we've been investing across robotics and physical AI for nearly a decade — with our very earliest bets including embodied AI leader Wayve and lab automation company Automata. So what has changed in the recent cycle – if anything?

On one hand, we are seeing a convergence of structural forces: declining hardware costs, chronic labour shortages, rising wage pressure, renewed political appetite for reshoring, and breakthroughs in AI-driven autonomy. And as related to the latter, the scope of environments in which robots can operate is expanding: from tightly constrained industrial settings; to semi-structured environments like warehouses, labs, and construction sites; to unstructured environments such as homes, outdoor spaces, and dynamic industrial workflows (though we note progress here is still far from the promise of full autonomy).

Historically, robotics adoption has been limited by brittle automation, long deployment timelines, and high capex, creating a ceiling on where robots could provide ROI. That ceiling is now lifting due to several developments:

• learning-based autonomy is starting to generalise across tasks in ways that rule-based systems never could

• advances in sensing, actuation, and simulation are making it possible to handle real-world variability that would have stumped robots in the past

• business model innovation — Robot as a Service, software-first approaches — is lowering the adoption friction for customers and increasing recurring revenue predictability

These shifts create a large and expanding investable landscape, where the TAM is not defined by today’s robot penetration but by the vast frontier of tasks that remain unautomated. Most physical tasks in the global economy - material handling, inspection, fabrication, assembly, logistics, food prep, maintenance - are still done manually. Even modest penetration across these categories translates into billions in enterprise value.

The opportunity 

The European opportunity is particularly exciting here given the deep talent density, industrial DNA and maturing ecosystem across both founders and investors. This is further buoyed by strategic tailwinds related to sovereignty and reshoring.

Europe has deep robotics & embodied-AI talent density unmatched outside the US and Asia. European universities produce a disproportionate share of world-class robotics researchers, perception scientists, and embodied-AI specialists. Key hubs continue to grow around Zurich (ETH, EPFL), Munich (TUM), and London (Imperial, Oxford, Cambridge, UCL). Specific labs like the Robotics Systems Lab at ETH and Robot Learning Lab at Imperial are steadily generating a pipeline of exceptional founders. The DeepMind diaspora and networks forming around Palantir, 1X and others are increasingly starting and funding companies. 

Europe’s industrial DNA and proximity to customers in this segment lend an interesting advantage too. It is a global center for industrial automation, manufacturing and machine tools. Manufacturing, logistics, welding, machine tending, agriculture, and defence roles are already chronically understaffed. German Mittelstand, and European/UK Tier-1 OEMs provide early customers willing to co-develop robotic workflows. Corporate partners like Siemens, ABB, Bosch, Rolls-Royce, Airbus are increasingly open to piloting early-stage technologies and crucially customers are no longer only experimenting, but actually starting to buy. This creates the conditions for improved adoption (though uptake often still considerably lags that in the US).

A range of strategic tailwinds add to the momentum. Geopolitical and national-security considerations have increased demand for autonomy, drones, and supply-chain resilience. European governments are actively pushing for technological sovereignty in defence and manufacturing - and robotics feature centrally in this policy priority. Indeed, Europe has increasingly become a leader in drone innovation, as tight-iteration cycles and feedback from the Ukrainian frontline inform an ever-growing ecosystem of suppliers.

It is still early innings. We believe we’re currently within a key window, as Europe’s ability to form champions is strongest before the market consolidates around a small number of global autonomy providers. We are bullish on existing bets we have made across the space and continue to actively focus on the opportunity set here.

How we think about investing

When underwriting investments in this space, we leverage several different frameworks

1) Team

We look to back teams who understand:

• where rapid iteration and customer learning cycles are possible

• how to design for deployment in real environments rather than demos (and have experience doing so with actual customers)

• how to combine hardware and software into defensible systems

• how to build business models that compound over time

This is not one-size-fits all, as certain approaches will justify more research heavy teams at time of investment at pre-seed (for example, where the Series A may be entirely science and/or tech milestone-based) whereas for others, customer and GTM will be more paramount from day 1 (for example, where there is little IP).

2) Business model and value accrual

We also work backwards from where companies in the space may ultimately trade on public markets, based on business model and company identity.

Is the company likely to be valued as a proxy for premium labor vs. a fundamentally recurring revenue generating software business vs. a hardware business (categorised by one-time sales and no recurring revenue)?

• RaaS: Tasks replacing labor (cleaning, picking, patrolling, QSR, agriculture), sell by the hour / task

• Vertically integrated: Software+hardware+workflow, sell full stack

• Foundational model: Companies building autonomy+intelligence layer, sell via licensing

Of course there will be overlap across these segments (ex: foundational model companies going vertically integrated, buying their own factories or building their own hardware). And at pre-seed often business model is still an open question of course!

3) Bottleneck / part of the stack

When it comes to achieving dexterous manipulation or scalable autonomy, consensus is lacking on where the ultimate unlock will come from, whether that be the tactile / hardware stack (biomimetic skins, force sensors, etc.), the model architecture (VLA, RL, imitation learning, etc.), or the data collection approach (teleoperation, self-supervised, synthetic data, sim-to-real, etc.). 

However, there seems to be some convergence around the data bottleneck issue: compared to LLMs or VLMs, data for robotics simply isn’t available at a similar order of magnitude. Robots require a lot of data to learn new tasks (the amount per task varies greatly based on the learning method and task itself), and the physical interaction data required for VLAs is limited. But to assume that if we simply had internet scale data the problem would be solved is an oversimplification.

Part of the challenge is that different approaches succeed in narrow domains but fail to generalise. For example:

• Vision-only manipulation may fail when contact dynamics matter

• Tactile-rich systems lack scalable datasets and are expensive

• Foundation-model control policies lack precision without good sensors

• Tele-op datasets are hard to scale beyond a few hand types or tasks

Key questions we double click on: What is off the shelf vs. what is being innovated on? Where will the IP sit, and how robust is it? How robust is the supply chain? Is this a marginal or step-change level innovation? Will the next round be based on science / tech / commercial milestones?

Where we are (and aren’t) investing

We believe that as autonomy requirements continue to rise, the ultimate unlock will likely be a co-evolution of software and hardware, not a single breakthrough in tactile sensing, foundation-model architecture, or data collection strategy. More recently we have invested in teams taking both horizontal approaches, like AMI and Mistral, Reimagine Robotics and more vertical approaches. With regards to the latter, we have invested across the following approaches:

• In tightly-scoped environments where autonomy requirements are more fixed and systems are more closed-loop, teams focused on high mix, low volume cases, where automation requires a high level of adaptability and flexibility (ie Rectify - manufacturing, Tekton - welding cells)

• In more unstructured, dynamic environments where autonomy requirements are more fluid, cultivating deep buy-in and goodwill with early partners is key as edge cases will inevitably flare up (ie Minerva - building the integrated hardware + autonomy stack for unmanned systems)

We’re excited about opportunities across a wide range of verticals including manufacturing, lab automation and pharma, energy and data centers, and more.

We tend to avoid pure component plays (e.g. sensors-only, tendons-only) that lack ability to control the end solution, drawing lessons from the autonomous driving sector. We do not believe the market will converge on a single winning form factor (e.g., humanoids vs quadrupeds vs arms). Instead, robotics adoption will follow a matrix of use case × environment, where different embodiments win in different contexts (e.g. quadrupeds for perimeter security in constrained environments vs. humanoids in the home, once dexterity is high enough). 

Learnings

 We continuously pressure test our views with learnings (of which there are many!) from our early robotics investments. A few we’ve been chewing on lately

1) Team composition: Don’t underestimate the value of deep industry operating experience. Also, hardware is hard

• The kinetic energy of star PhD engineering talent is often unlocked by deep industry operating experience. 

• Doing customer discovery without the experience of the inner workings of the industry can generate high risk of asking the wrong questions or not being able to challenge the insights being gathered.

• Don’t underestimate the value of someone on the team who understands in exceptional detail 1) the exact form factor and specs that would be required; 2) the human operations that would need to interact with the robotics; and 3) what it realistic to assume people will change in their core operations vs. what would be impossible to change. 

2) General vs. Applied: Each has their place, with the right team composition

• As related to above, investing in general robotics companies without an industry focus is high risk - especially if you plan to build foundational tech. The first 2-3 hypotheses around where to apply the tech are likely to be wrong, massively slowing down PMF

3) Rapid testing: Focus on where rapid learning from simple prototypes is possible

• If meaningful testing requires a fully integrated, complex full-stack system, the learning cycle is too slow, too costly, and too assumption-heavy

• If not feasible to break down to component builds that can be tested quickly, you are limited in how fast you can learn - you have to make too many assumptions upfront in an MVP which is complex and expensive before taking to market

4) Simulation vs prototype: Why not both?

• Simulations are helpful but not a panacea when it comes to how robots will actually behave in a live environment with customers

• Focus needs to be on speed-to-prototype without making too many assumptions from simulations 

5) Fundraising: Don’t underestimate capital requirements

• Series A rounds (when initial MVP's are validated) often will be bigger than initially thought to give proper runway for experimentations before any form of commercial scale is reached

• As a fund, be especially prepared to bridge where you have a) high conviction and b) a contrarian view 

6) Customer centricity: Never too early

• Build with a customer as soon as possible even if that means using commodity tech for a period of time as this allows you to refine the value proposition much more quickly way

If you are building in the space, we’d love to hear from you - reach out at adriana@firstminute.capital.

Adriana

P.S Thanks to all our founders and friends across the space that helped us think about our thesis here