Flexiona startup that “builds the brains of humanoid and human-capable robots” has raised $50 million in funding, Crunchbase News reports exclusively.
Founded in January by former Nvidia researchers, director general Nikita Rudin and Chief Technology Officer David Hoelleralong with Julian Nubert AND Fabian TischhauserZurich-based Flexion has now raised a total of $57.35 million in funding.
The company plans to use a part of its recent capital to open a U.S. headquarters in the Bay Area.
DST Global Partners, Nventures, Redalpin, Prosus ventures AND Moonfire Ventures participated in Series A financing.
Flexion was born from years of research in ETH Zurich and Rudin’s work at Nvidia.
The executive says most robots today rely on scripts, teleoperation or brittle task-specific code.
“It doesn’t scale,” he said.
According to Rudin, the Flexion platform replaces this with a full autonomy package, including language-based reasoning, motion generation through vision, language and motion, and transformer-based full-body control, “so robots can understand instructions, navigate the world, and adapt to new situations with minimal human involvement.”
He added: “Unlike companies that focus on a single robot package or narrow behaviors, our system is built to operate independently of morphology and tasks, making it a true general-purpose intelligence layer for robotics.”
Put simply, Flexion’s mission is to build the foundation of artificial intelligence “for the next era of robotics” by giving humanoid robots “the intelligence they need to transform industry and everyday life, making them safe, capable and irreplaceable partners of humans.”
The financing comes at a time that appears generally solid for robotics-related enterprise capital investment, with investments of over $10.7 billion worldwide as of November 19, according to Crunchbase data, already at the top of every full 12 months starting in 2021.
All sorts of work, every kind of tasks
Rudin said bend differs from other efforts involving basic robotics models in several respects.
First, it does not rely on manually developed behaviors or teleoperation, in which a human operator remotely controls and trains the robot. Instead, Flexion primarily uses synthetic data generated from high-performance physics simulations to train its models.
Second, it claims to use reinforcement learning techniques to deliver software that is “robust to the wide variety of the real world.”
This gives Flexion an advantage, Rudin says, because data generation is not limited by human labor and its models can “generalize and operate beyond the limitations” of a teleoperation setup.
The startup is initially focusing on humanoid and human-capable robots because, as Rudin puts it, they represent the highest-value opportunity, doing “useful work” in industrial applications, logistics, manufacturing and ultimately in areas corresponding to disaster response and planetary exploration.
Because the Flexion platform is morphology independent, the company also sees opportunities in wheeled platforms, multi-arm systems and other complex robotic forms. Over time, Flexion’s goal is to provide applications wherever robots need to perform long-term tasks on their very own.
Expansion plans
Currently, Flexion employs 31 employees. In addition to expanding into the United States, the company plans to use its recent capital to expand its Zurich-based research and development team, scale its fleets of computers and robots, and speed up the commercialization of its autonomous suite.
Rudin says the startup is already working with major OEM partners and the funding will help it expand that collaboration globally. Flexion licenses its software under an annual software licensing model per robot.
“There is a clear appetite for a purely software-based intelligence layer that can generalize actions to robotic bodies,” Rudin told Crunchbase News. But for now, the priority, he said, is to focus on developing core technologies.
Working at Nvidia gave Rudin a “deep appreciation” of the computational and data flywheel that enabled leaps in large language models.
“Working on basic robotic learning training tools gave me insight into the challenges robotics companies face: rebuilding the same components, acquiring the same knowledge, and fighting the same challenges,” he said.
“The hardest and most defendable part of the stack”
Filip Kneissinvestor in Redalpine, told Crunchbase News via email that after years of looking at the robotics space, Flexion stands out because, according to his company, it focuses on “the most difficult and defensible part of the stack: building a shared brain for robots.”
“They have already sent robots to work in the real world…” he said. “The ability to transform cutting-edge research into robust, field-proven independence is a big part of why we invested.”
Sandeep Bakshidirector of European investments at Prosus Ventures, said the start-up’s simulation approach was compelling because it had not seen any other robot modeler build with such an approach.
“Most players in today’s market use teleoperation-based solutions, which require hundreds of thousands of hours of manual human demonstrations. This is an approach that we believe is fundamentally unscalable in the long run,” he added. “Creators of robotic foundation models will eventually need to make extensive use of simulation-based training, and the Flexion team is best positioned to win with this approach.”