Swiss Startup Trains Humanoid Robot in Simulation to Retrieve Parcel Using Stairs and Elevator

Wired reported that Flexion Robotics, a Swiss startup founded by former Nvidia robotics researchers, has developed a system that trains humanoid robots to perform tasks such as opening doors, climbing stairs, and carrying boxes. The system teaches individual skills in simulation before a master AI model combines them to complete longer sequences.

A modified Unitree humanoid robot received the command to retrieve a delivered parcel of snacks using stairs and an elevator, then unpack the items and place them in a drawer.

The robot completed the sequence autonomously. Nikita Rudin, cofounder and CEO of Flexion Robotics and a former Nvidia research scientist, said the software’s “secret ingredient” is its extensive use of reinforcement learning. Each layer of the system, from the master model to simulation and motor control, applies this trial-and-error method.

The main AI model digests videos of humans performing tasks and matches those actions to skills learned in simulation. Flexion Robotics is collaborating with multiple robotics companies and testing the software across different humanoid platforms. George Chowdhury, an analyst at ABI Research, said the humanoid hardware itself is not the key advance and that progress depends on the underlying AI models.

ABI Research estimates the market for robot foundation models could reach $150 billion by 2036.