This humanoid robot is a terrifyingly competent office intern

Maybe humanoid robots be able to Being, danceAnd sometimes Kick peopleBut to become truly As humans, they will need to learn how to do all kinds of menial chores at work.

Flexion Robotics, a Swiss startup founded by ex-patsNvidia Robotics researchers believe he has the solution. The company has developed a training method Robots To perform complex tasks involving simple skills such as opening doors, climbing stairs, and carrying boxes. The key is to teach the robots individual skills in the simulation, and then have the main AI algorithm decide how to use them.

Most demonstration videos show human beings trained to do a specific task, such as folding shirts or loading shelves. Typically, this is done through teleoperation – a person behind the scenes controlling the robot’s movements. But this approach does not work reliably when the robot is placed in unfamiliar settings. Flexion says its system is different — and more efficient — because it trains its robots in simulation and with limited human instruction.

The video below shows the software in action: The modified Unitree robot operates autonomously after it receives the following command: “A package containing snacks has been delivered to Flexion. Retrieve it using the stairs and go up using the elevator. Then unpack it and place the items in the empty tray on the shelf in the snack area.”

Courtesy of Flexion

Flexion’s approach works by combining different AI systems.

The main AI model figures out how to do household chores by absorbing videos of humans doing different things. The software then matches the learned skills – captured in the simulation – to the videos and performs those tasks in the real world. In order to get to the mailroom in an office, for example, the model may have learned that it needs to open certain doors and use the elevator. The system also controls the robot’s motors, allowing it to walk, move its limbs and maintain balance.

According to Nikita Rudin, co-founder and CEO of Flexion and a former robotics research scientist at Nvidia, the software’s “secret ingredient” is its extensive use of reinforcement learning, which trains computers to master tasks through trial and error. Every layer of software, from the main AI model to simulation to motor control, uses this approach.

Courtesy of Flexion