Flexion Robotics, a Swiss startup founded by former Nvidia robotics researchers, demonstrated in 2026 a system that trains humanoid robots to complete multi-step office tasks without direct human control.
The company’s approach combines multiple AI systems to allow a robot to interpret a plain-language instruction and carry it out independently. In a demo video, a modified Unitree humanoid robot receives the command: “A parcel with snacks has been delivered for Flexion. 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 completes the task autonomously.
According to cofounder and CEO Nikita Rudin, a former robotics research scientist at Nvidia, the system’s core strength is its extensive use of reinforcement learning — a method that trains software to master tasks through trial and error. This approach runs through every layer of the system, from the master AI model to simulation to motor control.
The master AI model learns by processing videos of humans performing various activities. It then matches those observed behaviors to skills the robot has already practiced in simulation, and applies them in real-world environments. A separate layer handles motor control, managing the robot’s walking, limb movement, and balance.
Flexion says this differs from the common industry practice of teleoperation, where a human operator controls a robot’s movements during demonstrations. That method, the company argues, tends to break down when robots encounter unfamiliar settings. Flexion’s simulation-based training with limited human instruction is designed to be more reliable across varied environments.
The development suggests that humanoid robots may be moving closer to performing practical, multi-step workplace tasks without requiring constant human guidance — though the technology remains at the demonstration stage.
Source: WIRED