Nomagic's AI Robot Brain Achieves Success with Vision-Language-Action Model

Nomagic, an innovative company based in Warsaw, Poland, and Sandy Springs, Georgia, has made significant strides in the field of embodied artificial intelligence. The company’s approach to developing AI systems for robots is distinct from that of its competitors, focusing on creating models that excel at specific tasks right out of the box rather than general-purpose ‘robot brains.’ This strategy aims to eventually build towards a more comprehensive system by mastering individual tasks first.

The concept of embodied AI has gained traction in Silicon Valley, with investors betting big on systems that can interact with the physical world through robotic devices. Many startups are working on developing general-purpose AI models for robots, which would enable them to perform various tasks without extensive programming. However, these models often fall short of human-level accuracy and require significant task-specific training before achieving reliable results.

Nomagic’s approach is centered around creating highly accurate AI robot brains that can tackle specific challenges with ease. To pursue this goal, the company established an AI research lab led by Markus Wulfmeier, a former Google DeepMind robotics researcher who now serves as Nomagic’s chief scientist. This move marked a significant step towards developing more effective and efficient AI systems for robots.

Nomagic has recently announced that it has successfully deployed its first vision-language-action (VLA) model to paying customers. The VLA system is designed to perceive objects in the world, understand text-based instructions from humans, and take actions accordingly. This achievement makes Nomagic one of the pioneers in running VLAs in a live production environment rather than lab experiments or staged demos.

The early results of Nomagic’s VLA deployment are promising, with the company reporting that it has roughly halved the rate of robot-caused interventions in warehouse operations. The system is being used by Brack.Alltron, Switzerland’s second-largest e-commerce platform, which relies on robots from Nomagic to automate order picking and packing tasks.

Roland Brack, founder and owner of Brack.Alltron, expressed his enthusiasm for Nomagic’s VLA systems, stating that they have marked a significant step forward in the company’s automation efforts. According to Brack, the addition of these intelligent systems has enabled them to run autonomous shifts through nights and Sundays without increasing pressure on their human workforce.

Despite its success, Nomagic acknowledges that its VLA system is not yet perfect, with an accuracy rate below 99.9% in specific tasks. However, the company has developed a system around the VLA by integrating it with older ‘classical’ robotics software that acts as a safety net and error catcher. This approach ensures that the entire system can be trusted to operate reliably in customer warehouses.

Kacper Nowicki, Nomagic’s co-founder and CEO, explained that achieving high reliability is crucial for companies operating in the physical world. He emphasized that 99.9% accuracy is not just a marketing number but rather the minimum required to gain entry into most facilities. To address this challenge, Nomagic has developed a harness system that complements its VLA model, allowing it to operate safely and efficiently from day one.

Nomagic’s approach stands in contrast to many of its competitors, which focus on developing general-purpose AI models for robots. Markus Wulfmeier, chief scientist at Nomagic, highlighted the limitations of this strategy, stating that most companies are racing to build highly versatile robot brains without considering the practical challenges of mastering specific tasks.

Wulfmeier pointed out that the physical world is dominated by a long tail of rare situations, making it difficult for AI models to achieve high accuracy. He emphasized that relying on simulation or human teleoperation alone cannot economically close the remaining gap to the level required in real-world settings. Instead, Nomagic trains its VLA model using data from its existing fleet of robots deployed with customers.

Nomagic’s unique advantage lies in its ability to gather vast amounts of real-world data from its operational robots. This data is used to train and refine the VLA model, which Wulfmeier describes as unusually rich and diverse. By leveraging this approach, Nomagic aims to develop more effective AI systems for robots that can automate tasks with greater accuracy.

Tristan d’Orgeval, co-founder and chief strategy officer at Nomagic, stressed the importance of deploying robots in real-world settings first rather than relying on lab experiments or simulations. He noted that this approach allows companies like Nomagic to develop capable AI systems that emerge from practical experience rather than theoretical models.

The company’s focus on automation has earned it recognition within the industry. Recently, Nomagic won the 2026 International Intralogistics and Forklift Truck of the Year (IFOY) Award for its Shoebox Picker device. This achievement highlights the potential of Nomagic’s AI systems to tackle complex challenges in warehouse automation.

Nomagic’s approach marks a significant departure from traditional methods used by many companies working on embodied AI. By focusing on developing highly accurate models that excel at specific tasks, Nomagic aims to create more effective and efficient AI systems for robots.