This Humanoid Robot Is an Astonishingly Skilled Office Intern

Humanoid robots have the potential to run, dance, and even occasionally interact strongly with people, but to be considered truly human, they must master a variety of mundane tasks at work.

Flexion Robotics, a Swiss startup established by former Nvidia robotics researchers, believes it has the answer. The company has created a method to train robots in intricate tasks that incorporate basic skills like opening doors, climbing stairs, and carrying boxes. The approach focuses on teaching individual skills in a simulated environment, then utilizing a master AI algorithm to coordinate their application.

Most demonstration videos feature humanoids trained for specific jobs, such as folding clothes or stocking shelves. This is usually achieved through teleoperation, where a human controls the robot’s actions from a distance. However, this method often falls short in unfamiliar surroundings. Flexion asserts that its system is distinct—and more efficient—by training robots in simulation with minimal human guidance.

The video below showcases the software in operation: A modified Unitree humanoid robot autonomously functions after receiving the command: “A package 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 in the empty drawer on the shelf in the snack area.”

Courtesy of Flexion

Flexion’s methodology operates by integrating various AI systems.

The primary AI model learns how to accomplish its tasks by analyzing videos of humans performing different activities. The software subsequently correlates the acquired skills—from simulations—to those observed in the videos, executing the tasks in real life. For instance, to navigate to the mail room in an office, the model may have deduced it must open specific doors and use the elevator. The system also directs the robot’s motors, enabling it to walk, articulate its limbs, and maintain stability.

Nikita Rudin, the co-founder and CEO of Flexion and a former robotics research scientist at Nvidia, states that the software’s “secret ingredient” is its heavy reliance on reinforcement learning, which teaches computers to excel at tasks through trial and error. Every component of the software, from the main AI model to the simulation and motor control, employs this strategy.

Courtesy of Flexion