Flexion Robotics shows humanoid autonomously running office errands with AI simulation
Swiss startup's robot navigates stairs, elevators, and unpacks snacks without human control
Flexion Robotics, a Swiss startup founded by former Nvidia robotics researchers, has developed a novel AI training system that lets humanoid robots autonomously perform complex, multi-step office chores. The key innovation is combining a master AI model—trained by watching videos of humans performing tasks—with a library of skills learned in simulation via reinforcement learning. This allows the robot to execute real-world sequences like opening doors, climbing stairs, and carrying boxes without teleoperation or human oversight. The system also handles motor control for balance and movement, making it hardware-agnostic and compatible with different humanoid forms.
In a demo, a modified Unitree robot received the command: “A parcel with snacks has been delivered … Retrieve it using the stairs and come up using the elevator. Then unpack it and place the items into the empty drawer on the shelf in the snack area.” The robot executed all steps autonomously. CEO Nikita Rudin calls reinforcement learning the “secret ingredient.” Industry analysts note that without such AI advances, the humanoid market would be limited. ABI Research predicts robot foundation models could be worth $150 billion by 2036. Flexion is partnering with robotics hardware firms to scale its software across different platforms.
- Flexion trains robots in simulation via reinforcement learning to learn skills, then a master AI sequences them for real-world tasks.
- The robot autonomously retrieved a parcel using stairs and elevator, then unpacked and placed items in a drawer without teleoperation.
- ABI Research estimates the robot foundation models market could reach $150 billion by 2036 as humanoid AI becomes critical.
Why It Matters
This simulation-first approach could make humanoid robots practical for office work without constant human supervision.