Falkenlab closed an $88 million Series B on Wednesday led by Lightspeed Venture Partners, with participation from Atomico, Swisscom Ventures, and the venture arm of Migros. The Zurich company, spun out of ETH Zurich's robotics group in 2021, builds picking arms for grocery and pharmaceutical distribution centers. Its distinguishing claim is commissioning time. A conventional picking cell takes eight to twelve weeks to configure for a new facility and product mix. Falkenlab says its system reaches production accuracy in under 30 hours using a combination of teleoperation data collection and a policy that adapts on site. Three European grocery chains are running it in production across seven facilities.
The commissioning problem
Warehouse robotics has a dirty secret: the robot is cheap relative to the integration. A picking cell costs perhaps EUR 180,000 in hardware and another EUR 300,000 to EUR 500,000 in engineering to make it work in a specific building with specific products, specific totes, and specific lighting. That integration cost is what keeps automation out of mid-size distribution centers, and it recurs partially whenever the product assortment changes, which in grocery happens continuously.
Falkenlab attacks the integration cost rather than the hardware cost. Its arms are conventional, sourced from a Japanese supplier with a custom gripper. The software is the product: a vision and manipulation policy pretrained on roughly 40,000 hours of picking data across the company's deployments, which adapts to a new facility through a short teleoperation session where human operators handle the items the model is uncertain about, generating exactly the data the model needs.
How the on-site learning actually works
On day one, the system runs at low speed with a human teleoperator supervising through a workstation that can be in the building or in Falkenlab's Zurich operations center. The policy attempts each pick and hands off when its confidence falls below a threshold, which initially happens on perhaps 30 percent of items. Each takeover generates a labeled demonstration. Overnight the company fine-tunes the site-specific policy adapter on that day's data, and by the second shift the takeover rate typically falls under 5 percent.
Chief executive Nora Steiner, who did her doctorate on manipulation under uncertainty, was careful to describe the limits. The system handles rigid and semi-rigid items well, deformable packaging adequately, and loose produce poorly. Bags of salad remain unsolved and Falkenlab does not pretend otherwise. Facilities running the system route those items to human stations, which means the automation covers roughly 78 percent of picks at the grocery customers rather than the whole operation.
The story is rarely the launch. It is what breaks, what ships, and who owns the mess at 2 a.m.
The economics customers are seeing
A Swiss grocery customer shared numbers on condition its name not appear. It runs eleven Falkenlab cells across two facilities, at a total cost including hardware, integration, and the annual software subscription of about EUR 3.1 million over the first two years. The cells replaced roughly 34 full-time-equivalent picking positions, though the company says it redeployed rather than eliminated most of those roles because it cannot hire warehouse staff at any price in the region. Payback on the deployment lands around 26 months.
That is a decent but not spectacular return, and it is heavily dependent on labor cost. In Switzerland, where warehouse wages run above EUR 62,000 fully loaded, the math works. In Poland or Romania it does not, which constrains where Falkenlab can sell in Europe. The company's answer is that labor costs are converging upward across the continent and that the calculation will work in more markets each year, which is a reasonable bet and also a bet on macroeconomic conditions rather than on product.
The competition
Covariant, now largely absorbed into Amazon, pioneered the learned-policy approach and its people are inside the largest logistics operator on earth. Ambi Robotics and Osaro compete in North America. Berkshire Grey, after a difficult few years, remains a factor in large deployments. In Europe specifically the field is thinner, with Magazino in Munich focused on a different form factor and Dexory doing inventory rather than picking.
Falkenlab's geographic position is genuinely useful. European grocers are wary of American robotics vendors for data residency reasons and because support response time across an ocean is a real operational cost. Lightspeed partner Rytis Vitkauskas, who joins the board, said the European focus was central to the thesis and that the company should resist expanding to the United States before it saturates its home market, which is not advice venture investors usually give.
Risks and the road ahead
The technical risk is that on-site adaptation stops improving. The company's data advantage compounds only if new deployments teach the base model things it did not know, and there is a plausible ceiling where additional grocery warehouses add nothing. Steiner acknowledged this and said the response is expanding into pharmaceutical distribution and industrial parts, both of which present different manipulation challenges and both of which are in pilot.
The commercial risk is customer concentration. Three customers account for 84 percent of revenue, and one of them is an investor in this round, which is a structure that flatters current numbers and creates exposure. The Series B funds a sales team outside Switzerland for the first time, growing headcount from 62 to about 110 by the end of 2027, with the explicit goal of getting the top customer below half of revenue before the next round.
Skarvonix will keep following this beat with reporting grounded in how systems behave outside the launch keynote.
- Funding
