Robotics
Swiss Startup Teaches Humanoid Robots to Master Office Tasks Through Simulation and AI
Flexion Robotics has cracked a key challenge: training humanoids to autonomously handle complex workplace duties by combining AI models with reinforcement learning.
WIRED — AI
Flexion Robotics, a Swiss company founded by former Nvidia researchers, has developed a training method that enables humanoid robots to perform intricate workplace tasks—from retrieving packages to stocking shelves—without constant human remote control.
How the Technology Works#
The system operates by layering multiple AI approaches. A primary AI model learns from videos of humans performing various actions, then maps those learned skills—developed through simulated training—onto real-world tasks. The software orchestrates everything the robot physically needs to do: walking, moving limbs, balancing, and controlling its motors.
The company demonstrated the approach with a modified Unitree humanoid that, when given a multi-step instruction like "retrieve a parcel from the mail room using stairs and elevator, then stock items in the snack area," executed the entire sequence autonomously.
According to CEO Nikita Rudin, reinforcement learning—a technique where systems improve through trial-and-error—serves as the foundation for each component of the software stack. "This approach makes the system more robust when robots encounter unfamiliar environments, unlike robots trained solely through teleoperation," Rudin explained.
Why This Matters for the Market#
Industry analysts view the breakthrough as potentially transformative. George Chowdhury, a roboticist analyst, notes that "the AI models backing humanoids are what truly matter, not the hardware itself." Market researchers estimate the robot foundation model sector could reach $150 billion by 2036.
Tech executives including Elon Musk and Jensen Huang have long predicted humanoids could displace significant portions of the workforce. Yet realizing that potential depends on solving the autonomy problem that Flexion is tackling.
Challenges Ahead#
Flexion plans to collaborate with multiple robotics manufacturers and position itself as hardware-agnostic. Still, Chowdhury cautions that the startup faces intense competition and must work closely with hardware partners to commercialize its software. Without a reliable way to program humanoids as Flexion demonstrates, he argues, "the market fundamentally doesn't exist yet."
