Flexion's AI Trains Humanoids to Run Office Errands
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Flexion Robotics, a Swiss startup founded by ex-Nvidia researchers, developed AI software enabling humanoid robots to autonomously perform multi-step office tasks like opening doors, climbing stairs, and carrying boxes
- The system trains individual skills in simulation using reinforcement learning, then has a master AI algorithm combine them — sidestepping the unreliability of teleoperation-based training used for single-task demos like shirt-folding
- Nikita Rudin, Flexion's CEO and former Nvidia robotics research scientist, called reinforcement learning the software's 'secret ingredient,' noting every layer from the master model to motor control uses it
- A modified Unitree humanoid robot autonomously executed a single text command to retrieve a snack parcel via stairs and elevator and stock it in a drawer
- ABI Research estimates the market for robot foundation models could reach $150 billion by 2036, with analyst George Chowdhury warning Flexion must partner with hardware makers or 'there isn't really a market here'
Why it matters: The humanoid robot race is being framed as a hardware contest, but Flexion's demo makes the software layer the gating factor: analyst George Chowdhury says without programmable AI that handles novel environments, 'there isn't really a market here.' ABI Research pegs that robot foundation model market at $150 billion by 2036 — and Flexion's hardware-agnostic stance gives its software potential to run across multiple robot bodies.
