This Humanoid Robot Is A Very Skilled Worker

Humanoid robots are gone to be able to run, danceand sometimes to push peoplebut to be honestly people, they will have to learn to do all kinds of menial jobs at work.

Flexion Robotics, a Swiss startup founded by ex-Nvidia robotics researchers, think it has an answer. The company has developed a training system robots performing complex tasks that involve simple skills such as opening doors, climbing stairs, and carrying boxes. The key is to teach robots individual skills, and then have a master AI algorithm figure out how to use them.

Many animations feature humanoids that have been trained to perform a specific task, such as folding shirts or lifting shelves. In most cases, this is done through a telepath – an invisible person who controls the robot’s movements. But this method does not work reliably when the robot is placed in an unfamiliar environment. Flexion says its system is different—and it works—because it trains its robots asynchronously and without human instruction.

The video below shows the program in action: The modified Unitree humanoid robot performs an autonomous task after receiving the following command: “A program with snacks is given to Flexion. Bring it down using the stairs and go up using the elevator.”

Courtesy of Flexion

The Flexion method works by combining different AI systems.

The main type of AI shows how to do its homework by digesting videos of people doing different things. The program combines the skills learned – which have been learned through simulation – with videos and doing those tasks in the real world. To get to the mail room in the office, for example, the model must have learned that he must open certain doors and use the elevator. The system also controls the robot’s motors, allowing it to walk, move its legs, and stabilize itself.

According to Nikita Rudin, cofounder and CEO of Flexion and a former robotics research scientist at Nvidia, the “secret” of the program is the extensive use of reinforcement learning, which teaches computers to master tasks through trial and error. Every part of the software, from the AI ​​master model to the simulation and driving, uses this method.

Courtesy of Flexion