Nomagic uses production data to bridge the reliability gap in warehouse robotics
The Polish startup is deploying a 'production-first' data strategy to reduce robot interventions and achieve industrial-grade reliability.
Polish robotics startup Nomagic is positioning real-world production data as its primary competitive advantage in the race to automate warehouses. By deploying functional robots first and training AI on the resulting data, the company aims to solve the persistent reliability issues that plague embodied AI.
Nomagic recently deployed its first vision-language-action (VLA) model to paying customers, specifically targeting the complex 'edge cases' that typically trigger human intervention. This VLA model has already roughly halved the rate of robot-caused interventions in live operations for customers such as Brack.Alltron. To maintain industrial standards during this iterative AI rollout, the company employs a 'harness' of classical robotics software designed to ensure 99.9% reliability.
The Data Moat
Historically, warehouse automation has struggled with the 'long tail' of rare errors. Many AI labs rely on 'sim-to-real' training—using simulations to teach robots—or human teleoperation. However, these methods often plateau at approximately 80% accuracy, a level insufficient for the economic demands of industrial logistics.
Nomagic is countering this by building a massive dataset from active deployments. Its fleet currently generates millions of picks every month, including two million monthly picks with Zalando alone. This volume of real-world interaction provides a rich training set that simulation cannot replicate. The company is betting that mastery must be earned in real deployments first.
Industrial Implications
This shift from general-purpose AI research to a 'production-first' approach represents a strategic pivot in the field of embodied AI. For Nomagic, the goal is not to create a universal robot brain, but to achieve task-specific mastery through scale. If the company can prove that live production data is the only viable path to move from 80% to 99.9% reliability, it creates a significant barrier to entry for research-heavy competitors who lack large-scale physical deployments.
CEO Kacper Nowicki has emphasized the necessity of this precision, noting that 99.9% reliability is not a marketing number, but the cost of being allowed in the building.
Scaling the Vision
To fuel this expansion, Nomagic raised $44 million in a Series B funding round led by the EBRD, with participation from Almaz Capital and Khosla Ventures. The capital supports the company's effort to scale its fleet and further refine its VLA models.
Industry observers are now watching to see if Nomagic's data-centric approach can be scaled across different warehouse environments and product types. The primary remaining question is whether this 'production-first' moat can withstand the arrival of larger general-purpose models or if the specificity of warehouse data will remain the ultimate deciding factor in industrial viability.