Supply chain / Warehouse robotics · Deep dive

Sereact

A Stuttgart spin-out selling a vision-language-action 'brain' — Cortex, plus its PickGPT lineage — that runs on any warehouse or manufacturing robot arm, pitched as the software layer under BMW, PepsiCo and Zalando picking cells rather than yet another custom cell.

emerging

The question that decides it: Does Sereact's Cortex vision-language-action model actually generalize across SKUs, robot embodiments and customer sites without customer-specific fine-tuning — so a new arm goes live in a shift on subscription economics — or does every deployment slide into a services engagement (integration, cell design, per-customer retraining) that caps unit economics and lets a US-scale foundation-model player like Physical Intelligence or Skild, or the reverse-acquihired Covariant stack inside Amazon, out-generalize it once they turn their models loose on warehouses?

My take ★★ Very interesting ★ Interesting − Maybe ✗ Not interesting

Snapshot

Sereact is a five-year-old Stuttgart AI-robotics company selling a general-purpose brain — not a robot — for industrial manipulation. Its Cortex vision-language-action foundation model, and the PickGPT lineage that preceded it, is pitched under the tagline “One Brain. Any Robot.” to run picking cells, dual-arm returns stations and humanoid platforms on top of arms and grippers the customer already owns. By mid-2026 it says more than 200 systems are live across Europe and cumulative production picks have passed one billion, with BMW, Daimler Truck, Mercedes-Benz, PepsiCo, Active Ants, bol., Austrian Post and Rohlik Group as named accounts. In April 2026 it raised a $110M Series B led by Headline, which Zalando extended to $116M in July when it joined as a strategic investor to underwrite dual-arm returns processing. Sereact has opened a Charlestown, Massachusetts headquarters and started hiring the US commercial and application team the round was meant to fund.

Founding story

Ralf Gulde and Marc Tuscher met young and spent close to a decade in the same building. Both trained at the University of Stuttgart — Gulde in mechatronics, Tuscher in Informatik — and both did research from 2011 onward at ISW Stuttgart, the university’s institute for machine-tool control and production units, working on reinforcement learning and computer vision for manipulation. Both did applied stints at Festo, the Baden-Württemberg pneumatics and automation company that quietly employs a large share of the region’s collaborative-robot engineers. In interviews Gulde returns to a single conviction: the physics of picking objects out of a bin at industrial reliability is a first-principles problem, and the pattern of academic robotics research staying stuck in labs while European DCs and factories screamed for labour was overdue for a company that shipped rather than published.

They incorporated Sereact in 2021 out of Stuttgart’s Cyber Valley cluster, and the pitch from the start was software rather than hardware — a policy-and-perception layer that treats robot arms as commodities to be actuated by a smarter brain. Early 2023 produced PickGPT, the first commercially deployed vision-language-action model for logistics, which won the Robot Report’s RBR50 Innovation Award for 2024. That framing has driven every strategic choice since: hardware-agnostic integration first, foundation-model scale second, US expansion third. The prompt suggests a third co-founder named Michael Neumann; no public record ties that name to Sereact’s founding team, and every funding announcement lists only Gulde and Tuscher — a research gap worth flagging.

How it works

Physically, a Sereact system looks like an ordinary industrial picking cell. A robot arm — Sereact runs on a range of commercial platforms rather than shipping its own, though the company has not publicly enumerated the specific OEMs — sits over a conveyor or tote-induction point, wearing a suction-based or multi-finger gripper and a 2D/3D camera pair mounted on the wrist or above the cell. What differs is the software behind it. The PickGPT / Cortex stack ingests the camera stream, matches operator instructions (which can be typed as natural language rather than task IDs), reasons about what to grasp and in what order, and issues motion commands to the arm’s controller. Sereact’s own product literature describes zero-shot generalization to unseen SKUs — the model recognises an object it has never been trained on from a text description or a reference image, and picks it without a hand-tuned policy.

Cortex 2.0, announced in April 2026 and funded by the Series B, layers a world model on top of the VLA. Before executing a motion the system simulates candidate trajectories, scores each for physical outcome (grasp stability, collision risk, cycle time), and only then commits. The company frames the shift as moving “from try-and-see to plan-and-try,” planning in a learned visual latent space rather than in robot-specific joint coordinates so the same brain can be re-targeted from a single-arm picking cell to a dual-arm returns station to a humanoid without rewriting the policy. Beyond picking the same stack also runs Sereact Lens, a 3D perception layer for inventory and quality control, and the dual-arm returns cell Zalando is underwriting — a station that takes in a sealed returned parcel, opens it, extracts the garment, refolds it and reinducts it without human touch.

Product and business overview

Sereact sells three things that are really one thing marketed to different buyers. The primary product is the Cortex platform: the software brain itself, sold to warehouse and manufacturing operators as the intelligence layer under whatever manipulation hardware a facility already runs or intends to buy. Sitting on top of Cortex are application templates — pick-and-place, kitting (aimed at automotive), returns handling (aimed at apparel and e-commerce), palletising, and quality inspection — each of them a configured deployment of the same underlying model rather than a separately engineered product. Sereact Lens is the perception-only SKU for customers who want the vision and inventory stack without the manipulation piece. Cortex 2.0 is the model-family upgrade the Series B pays for: a VLA-plus-world-model architecture the company argues is the first that can safely reason about contact-rich tasks like assembly.

The commercial pitch to enterprise buyers is that this is the layer that makes hardware disposable. Because Cortex is embodiment-agnostic, a customer who has bought Sereact once can, in principle, re-target the same subscription to a new arm, a new gripper or eventually a humanoid without a services rebuild. That is the durability claim the whole valuation rests on. The counter-claim, which the company has not yet publicly refuted with independent unit-economic data, is that every real deployment still requires site-specific integration, camera calibration, gripper selection, safety fencing and enough teleoperation data to calibrate the model to a facility’s SKU tail — i.e. that Cortex is genuinely portable in the abstract but arrives in the field as a services engagement.

Business model and pricing

Sereact publishes no pricing and has not disclosed how revenue is booked. Public commentary describes tailored project proposals sized to a customer’s cell count and SKU set — which is consistent with a hybrid: an upfront integration or hardware fee, plus a recurring per-cell or per-arm software subscription, plus, on the returns and kitting products, a possible per-cycle or per-unit performance component. The company argues its systems run about 30% cheaper than the equivalent manual station (company statement, Jul 2026), payback in weeks not years, and it will set up a cell in “one day” — the Series B narrative in every trade press write-up. None of those numbers is audited.

The strategic-investor mix around the Series B is doing real commercial work. Zalando’s July 2026 check ties the returns-processing product to the buyer with the loudest returns problem in European e-commerce (returns account for roughly half of everything Zalando ships, per its own disclosures). Bullhound and Felix Capital bring European growth relationships; Daphni is the French angle for the French auto and retail customers Sereact already sells into; Headline anchors the US expansion with San Francisco and Berlin desks. Latka’s public directory lists ~$6.4M revenue on a 68-person team for 2025 — unverified and directional, but if roughly right it means Sereact went into the Series B with revenue-per-head of under $100K, which is consistent with early enterprise pilots rather than a scaled SaaS-like book.

Traction over time

By July 2026 Sereact’s own disclosures put the installed base at 200+ live systems across Europe and cumulative production picks past one billion — the largest AI-picking footprint any European vendor has publicly claimed. The named customer list has expanded from four in the 2024 Series A materials to eleven by the 2026 Series B (Active Ants, Austrian Post, BMW, bol., Daimler Truck, DeltiLog, Mercedes-Benz, Monta, MS Direct, PepsiCo, Rohlik Group), plus the strategic Zalando returns rollout. Employee count — 93 as of June 2026 per Revelio, up from 68 at end of 2025 — is small for the ambition, and the Boston hiring plan (10 starting, targeting 60 within 12 months) is where the money is going first.

Market analysis

The numbers depend on where the line is drawn. MarketsandMarkets sizes the broader warehouse-automation market at $27.4B in 2026 rising to ~$59.5B by 2030 (18.7% CAGR). The narrower warehouse-robotics segment is quoted at $7.35B in 2026 rising to $25.41B by 2034 by Fortune Business Insights (16.8% CAGR), and $10.96B rising to $24.55B by 2031 by Mordor Intelligence (17.5%). The bit Sereact really monetises — physical-AI software as a layer inside industrial robotics — is measured differently. MarketsandMarkets pegs the physical-AI category at $1.5B in 2026 to $15.24B by 2032 (47.2% CAGR), with logistics and supply chain the largest slice. The structural tailwind is the one every warehouse-robotics deck cites: European warehouses face labour costs 30-50% above US medians in the highest-wage countries, chronic staffing shortages that got worse post-pandemic, and rising minimum wages across Germany and the Netherlands that push the manual-cost baseline up every year. That is what makes a 30%-cheaper picking cell a live conversation at BMW, Zalando or Rohlik rather than a proof-of-concept discussion.

The uglier structural fact is that the same tailwinds apply to every US foundation-model rival, and their capital pools are an order of magnitude bigger. Sereact’s Series B is priced against Physical Intelligence at ~$5.6B (Nov 2025) and Skild AI at ~$15B (Mar 2026), neither of which sells warehouse manipulation today but both of which train the same class of model at 5-10x the scale.

Competitive intel

Four camps to hold in mind. The reverse-acquihired ghost, Covariant, is the reference case: Amazon licensed the models and took the founders in Aug 2024, leaving the market to argue whether best-in-class AI-picking IP inevitably ends inside a hyperscaler-operator. Every AI-picking Series B priced since is priced against that outcome. The direct US operators — Dexterity, Ambi Robotics, RightHand Robotics — compete on the North American 3PL and shipper accounts Sereact is now hiring against out of Boston, with earlier US channels (Dexterity’s Kawasaki tie-up, Ambi’s USPS and Fortune 500 shipper base, RightHand’s e-commerce installed base) and reasonably deep capital. The foundation-model bigs — Physical Intelligence, Skild — are the shadow rivals whose models are, by scale of training data and compute, harder to out-generalize; the bet that keeps Sereact independent is that vertical, hands-in-the-warehouse deployment data at 1B+ picks is a compounding moat those horizontal players do not yet have. The classic industrial incumbents — KUKA, ABB, Fanuc, plus vision-software specialists Fizyr and Euclid Labs — set the baseline: they own the arm-and-integrator channel and can bundle better vision at their own price point if a large enough customer forces it.

Symbotic and the ASRS incumbents are adjacent rather than direct: they sell a different physical envelope (case-handling and pallet-flow), but every dollar a customer spends on a full ASRS is a dollar not spent on a Sereact picking cell.

History and evolution

• 2021 — Ralf Gulde and Marc Tuscher spin Sereact out of ISW / University of Stuttgart into the Cyber Valley ecosystem.
• Early 2023 — PickGPT ships as an LLM-plus-vision robotics transformer; commercial deployments begin at European DCs.
• Aug 2023 — Seed round of $5M closes, led by Point Nine and Air Street Capital, with angels including Oliver Cameron and the ProGlove founders.
• 2024 — Wins RBR50 Robotics Innovation Award; installed base and named customers grow through European logistics and automotive (BMW, Daimler Truck, bol., Active Ants).
• Jan 2025 — €25M Series A led by Creandum; use of proceeds explicitly earmarked for US expansion and R&D on mobile-manipulation and humanoid platforms.
• Sep 2025 — Cortex introduced as the umbrella foundation-model brand for the picking, kitting, returns and inspection products.
• Apr 2026 — $110M Series B led by Headline; Cortex 2.0 (VLA + world model) launched; company opens its US HQ in Charlestown, near MassRobotics.
• Jul 2026 — Zalando joins the Series B as strategic investor, taking the round to $116M; total funding to date ~$146M.

What people say

The case for. European trade press — Tech.eu, Sifted, EU-Startups, MassRobotics, The Robot Report, SiliconANGLE — has consistently treated Sereact as the most deployed AI-picking vendor in Europe and the credible non-US answer to Covariant’s absorption into Amazon. The Bullhound Capital write-up frames PickGPT as a genuine engineering breakthrough in generalisation; Creandum’s Series A thesis argues that the ISW pedigree and the choice to lead with software rather than hardware produce a defensible moat as the installed base compounds picking data. Enterprise references are unusually senior for a company this size — BMW and Daimler Truck signalling that Sereact clears procurement bars at German industrial giants, PepsiCo signalling that the pitch travels to US-headquartered consumer goods, and Zalando’s July 2026 strategic check being the loudest possible endorsement from Europe’s biggest fashion e-commerce operator that the dual-arm returns product is worth backing. Company-cited unit economics — ~30% cheaper than manual, payback in weeks, one-day setup — are the numbers every sales conversation runs on if they hold.

The complaints. The disclosure gap is real. Sereact publishes no pricing, no per-customer ROI case study a third party has audited, no gross margin, no ARR, no live-cell count per customer and no throughput or first-attempt success-rate figures — the standard KPIs every industrial-buyer diligence deck asks about. The Latka-reported $6.4M 2025 revenue on 68 heads is directional but small, and it implies most of the 200+ live systems are either low-priced pilots or on-going services engagements rather than high-margin subscriptions. The competitive shadow is uncomfortable: Physical Intelligence and Skild AI together have raised ~10x Sereact’s total capital to train the same class of model, and Dexterity, Ambi and RightHand each hold US installed bases Sereact has not yet displaced. Robotics-industry Reddit and Hacker News threads on the broader “AI-picking” category recur to two complaints — that generalisation to unseen SKUs collapses on long-tail edge cases in real DCs, and that vendors’ hardware-agnostic claims paper over site-specific integration cost. And there is the Covariant precedent: the last European (well, Berkeley-founded) AI-picking company priced as an independent winner ended as an Amazon license.

Outlook: the open question

Sereact has more deployed AI-picking systems than any other named European vendor, a genuinely differentiated architecture (VLA plus, as of April 2026, an added world model), a Series B lead who can help fund US expansion, and the loudest possible European strategic in Zalando. It is asking the right first-principles question: why should every warehouse cell be a bespoke integration when a single learned brain could run all of them?

The falsifiable question is not whether Cortex works in a demo — that is settled. The bull case requires three things to be visibly true within 18-24 months. (a) A published, independently verifiable throughput-per-cell and first-attempt success-rate curve across the 200+ live systems, showing that new SKUs go live with hours of on-site data rather than weeks of bespoke fine-tuning — the generalization claim landing in the field. (b) The Zalando dual-arm returns rollout scaling to hundreds of cells across Zalando’s European fulfilment network on the same subscription unit economics as picking, proving that Cortex genuinely re-targets across embodiments and tasks without a services rebuild. (c) At least one US Fortune 500 anchor customer (a 3PL, a grocer, a consumer-goods shipper) live on Cortex in the first 12 months out of Boston, at margins that survive the Physical Intelligence / Skild pricing gravity once those horizontal models turn to warehouses. The bear case is the exact mirror. Cell-level economics stay opaque because every deployment is really a services engagement. The dual-arm returns pilot at Zalando is slower and more custom than the deck says. The US installed base does not appear or appears at prices Dexterity and Ambi can undercut. And when Physical Intelligence, Skild or Amazon-Covariant point a horizontal foundation model at logistics manipulation, Sereact’s 1B-picks data moat proves narrower than assumed and it ends where Covariant did — as a talent-and-license transaction rather than an independent public company.

Sereact is the most credible European entrant in a category where the definition of “winning” has shifted from best product to biggest model. Whether it is a category or a positioning is what the next 18 months answer.

How a challenger would attack it

Publish the numbers Sereact won’t. Sereact discloses no pricing, no audited ROI, no throughput or first-attempt success-rate curves — and the Latka-reported $6.4M of 2025 revenue against 200+ “live systems” implies most cells are pilots or services engagements, not scaled subscriptions. A challenger attacks straight into that disclosure gap: publish per-cell throughput, first-attempt pick rates and payback data verified by a named customer, and offer performance-priced contracts — pay per successful pick, penalties for downtime — that force Sereact to either match transparency or concede its “one-day setup, payback in weeks” numbers are marketing. The second vector is capital asymmetry: Physical Intelligence ($5.6B) and Skild (~$15B) are training the same VLA class at 5-10x the compute; the sharpest attack is not another Stuttgart lab but a warehouse-vertical application layer built on one of those horizontal models, arriving with better generalization for the long-tail SKUs where Reddit-documented AI-picking failures cluster. Third, timing: Sereact’s Boston beachhead is ten people chasing sixty hires against Dexterity’s Kawasaki channel, Ambi’s Fortune 500 shipper base and RightHand’s installed accounts — a US incumbent that locks 3PL master agreements in the next 12 months makes Sereact’s US expansion a land war against entrenched channels with a sub-100-person company split across an ocean.

Same playbook, new buyer

Take the hardware-agnostic brain to cells that aren’t picking. Sereact’s own architecture claim — plan in a learned latent space, re-target across embodiments — points at buyers it isn’t resourced to chase. The nearest is manufacturing assembly and machine tending in the German Mittelstand: contact-rich tasks Cortex 2.0’s world model is explicitly built for, sold to mid-size factories (RobCo’s customer base) rather than DC operators, where the buyer wants a smarter brain on an arm it already owns and no US foundation-model player has a channel. Sereact gestures at this with its kitting product but its Series B narrative, Zalando tie-up and Boston hiring all commit it to warehouse logistics; a focused rival owns industrial assembly before Sereact circles back. Second: the returns-processing cell as a standalone franchise — returns are half of what Zalando ships, and every apparel e-tailer has the same problem, but Sereact’s dual-arm product is now strategically entangled with one retailer-investor; a neutral vendor sells the identical station to Zalando’s competitors (ASOS, Shein’s European ops, department-store e-commerce) who won’t fund their rival’s cap table. Third: Japan and Korea — acute labor shortage, dense robotics OEM ecosystems that want a software brain to differentiate their arms, and no European or US VLA vendor selling there; an OEM-licensed model business reaches those arms through the channel Sereact’s direct-sales model ignores.

Sources and further reading

• Sereact — Series B announcement (Apr 2026) (company post, Apr 2026)
• Tech.eu — Zalando joins Sereact’s $116M Series B (20 Jul 2026)
• The Robot Report — Sereact gets Series B funding to expand Cortex 2.0, enter US (Apr 2026)
• Sifted — Sereact taps Headline to lead $110M Series B, plots US expansion (Apr 2026)
• MassRobotics — Sereact opens Massachusetts headquarters in US expansion (2026)
• Automated Warehouse — Sereact raises $110M to scale Cortex 2 world model (Apr 2026)
• Tech.eu — Sereact’s €25M Series A raise boosts AI robotics zero-shot visual reasoning (20 Jan 2025)
• Air Street Press — Announcing Sereact’s €25M Series A (Jan 2025)
• The Robot Report — Sereact’s PickGPT lets users instruct robots with natural language (2023)
• Creandum — Backing Sereact: Building a Brain for Robots (Jan 2025)
• WWD / Sourcing Journal — Betting on the future of physical AI, Zalando invests in Sereact (Jul 2026)
• GeekWire — Amazon hires Covariant founders, inks AI licensing deal (2 Sep 2024)
• Sacra — Skild AI funding and analysis (2026)
• The Robot Report — Physical Intelligence raises $600M to advance robot foundation models (Nov 2025)
• MarketsandMarkets — Physical AI market to $15.24B by 2032 (2026)