Humanoid robots could power start, danceand occasionally kick peoplebut become Really people, will have to learn to do all kinds of tiny tasks at work.
Flexion Robotics, a Swiss startup founded by former Nvidia robotics researchers, thinks it has a solution. The company has developed a way to train robots to perform elaborate tasks that require straightforward skills, such as opening doors, climbing stairs and moving boxes. The key is to teach the robots individual simulation skills and then have the master AI algorithm determine how to operate them.
Most demonstration videos feature humanoids trained to perform specific tasks, such as folding shirts or loading shelves. This is usually done through teleoperation – a person behind the scenes who controls the robot’s movements. However, this approach does not work reliably when the robot is in unknown conditions. Flexion says its system is different – and more proficient – because it trains its robots in simulations and with narrow human instruction.
The video below shows the software in action: A modified Unitree humanoid robot operates autonomously after receiving the following command: “A package of snacks has been delivered to Flexion. Take the stairs and take the elevator to pick it up. Then unpack and place the items in an empty drawer on a shelf in the snack area.”
Courtesy of Flexion
Flexion’s approach is to combine different artificial intelligence systems.
The main AI model determines how to perform its duties by digesting videos of people performing various activities. The software then matches the acquired skills – acquired in the simulation – to the videos and performs these tasks in the real world. For example, to get to the office mailbox, a model might have been told she had to open a certain door and operate 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 scientist at Nvidia, the software’s “secret ingredient” is its extensive operate of reinforcement learning, which teaches computers to perform tasks through trial and error. Every layer of software, from the core AI model to simulation to engine control, uses this approach.
Courtesy of Flexion