Nomagic’s Warehouse Robots Get an AI Brain: Halving Human Interventions
Nomagic, a Warsaw-based warehouse robotics firm, has made a significant leap forward in robotic automation. They have integrated a vision-language-action (VLA) model into live customer operations, achieving remarkable results. According to the company, this technology has almost halved the need for human assistance from their robots.
A Polish Robotics Company Puts AI into Action
Fortune reports that Nomagic is among the first to deploy a VLA model in real production settings, not just staged demonstrations. This breakthrough comes as many robot labs are still in the demo phase, chasing the elusive "general robot brain."
An Innovative Approach: Mastery Before Generality
Nomagic takes a unique approach by prioritizing mastery over generality. They aim to develop models that excel at specific tasks before building a more general system. This strategy is led by Markus Wulfmeier, a former Google DeepMind researcher and core member of the Gemini Robotics team, who heads Nomagic’s AI research lab.
Overcoming Challenges in the Physical World
The physical world presents unique challenges, with countless rare situations that can throw off even the most advanced systems. This was a significant hurdle for self-driving cars. Wulfmeier argues that achieving 80% accuracy in simulations or remote control training is not enough for real-world applications like warehouses, where robots need human intervention only once an hour, making the economics unviable.
Nomagic’s Solution: A Harness for Perfection
Recognizing these challenges, Nomagic has developed a "harness" that combines their VLA with classical software. This hybrid approach ensures error catching and safety enforcement, achieving the high standards required in warehouses. Co-founder and CEO Kacper Nowicki emphasizes that this harness allows them to clear the bar of 99.9% success while the AI continues to improve.
Data Advantage: Millions of Successful Picks
Nomagic’s edge lies in its vast data collection. Their deployed fleet already processes millions of successful picks each month, with two million coming from fashion platform Zalando alone. This data is invaluable for training their VLAs and pushing the boundaries of robotic automation.