Zero-Shot AI Robots Pick 60,000+ SKUs Without Item Training

Zero-Shot Vision Eliminates the SKU Training Bottleneck

Swiss eCommerce logistics provider MS Direct has deployed warehouse robots that pick and pack orders from an inventory of more than 60,000 SKUs without requiring training on individual products—solving a fundamental barrier that has kept most warehouse picking operations manual.

The robots, built by Stuttgart-based Sereact, handle roughly 1,500 single-item orders daily at MS Direct's largest fulfillment center in Arbon, working through the night shift. The deployment paid for itself in approximately nine months, according to a report from MassRobotics, with Switzerland's high labor costs accelerating the return.

MS Direct moves more than 32 million items annually. Its Arbon facility runs on a Kardex AutoStore grid system, where robots retrieve storage bins and deliver them to picking stations. For years, the company couldn't automate the actual picking: the first robot system it tested broke down when confronted with the product variety. Picking remained manual at eight of nine AutoStore ports.

Why it matters

SKU complexity and packaging variance represent the primary obstacles preventing lights-out warehouse operations across the industry. Conventional robotic systems require product images, grip instructions, or manual testing before handling unfamiliar items—work that must repeat whenever retailers add merchandise or change packaging. In eCommerce fulfillment, that happens constantly. Zero-shot AI removes that requirement, making automation economically viable for operators handling diverse, changing inventories rather than just high-volume, standardized goods.

Real-Time Vision and Grip Selection

Sereact's system analyzes each object in real time, identifying shape, material, and color, then selects an appropriate grip without product-specific training, according to Packaging Journal. The software adjusts grip and movement dynamically if an item shifts during a pick. One in roughly 53,000 picks requires remote human intervention, Sereact CEO Ralf Gulde told MassRobotics.

The model trains on production data rather than simulations. Every successful pick, failure, and recovery is captured with synchronized observations, robot state, gripper force feedback, and outcome, then used to continuously update the model. More than 200 Sereact systems are live across Europe and have completed over 1 billion production picks for customers including BMW, Daimler Truck, PepsiCo, and Austrian Post, MassRobotics reported.

Sereact raised $110 million in a Series B round led by Headline in April to scale its next-generation Cortex model and open its first U.S. office in Boston, Bloomberg reported.

Night-Shift Automation Improves Asset Utilization

MS Direct's AutoStore system had run predominantly during the day, leaving infrastructure idle at night. The Sereact robot now fills that gap, handling picking throughout the night and raising output from equipment the company already owned without adding night-shift labor. Employees continue to handle orders requiring judgment or falling outside the robot's range.

The challenge extends beyond mid-sized operators. Amazon's Vulcan robot, deployed in Spokane and Hamburg, uses force feedback sensors because most commercial robots cannot reliably detect or adapt to unexpected contact with items. Amazon is investing €10 billion in its European fulfillment network, with Vulcan and other AI-powered warehouse systems central to that expansion, Reuters reported. Amazon's fulfillment network spans more than 200 facilities in Europe alone.

Most third-party logistics providers run a fraction of that volume across dozens of clients, each with different and changing product catalogs—making zero-shot picking capability particularly valuable for the broader market.

These details were first reported by PYMNTS.