Learn about the Chinese startup that’s using artificial intelligence — and a small army of workers — to train robots

The real question is how effective AgiBot’s algorithms are at teaching robots new tricks. Using reinforcement learning to teach a robot tasks that require improvisation generally requires a lot of training data, and studies show that it cannot be fully mastered within a simulation.

AgiBot speeds up the learning process by having a human operator guide the robot through a task, providing the foundation for it to then learn on its own. Before co-founding AgiBot, chief scientist Jianlan Luo conducted cutting-edge research at UC Berkeley, including project Which involves robots acquiring skills through reinforcement learning with a human in the loop. This system was shown performing tasks including placing components on the motherboard.

AgiBot’s learning software, called Real-World Reinforcement Learning, only needs about ten minutes to train the robot to do a new task, Feng says. Rapid learning is important because production lines often change from week to week, or even during the same production process, and robots that can quickly master a new step can adapt alongside human workers.

Training robots in this way requires a lot of human effort. AgiBot has Robotic learning center It pays people to operate robots to help AI models learn new skills. Demand for this type of robot training data is increasing, with some American companies Paying workers wages in places like India To do manual work that serves as training data.

Jeff Schneider, a roboticist at Carnegie Mellon University who works on reinforcement learning, says the AgiBot uses cutting-edge technologies and should be able to automate tasks with high reliability. Schneider adds that other robotics companies will likely get involved in using reinforcement learning for manufacturing tasks.

AgiBot is a rising star in China, where interest in combining artificial intelligence and robotics is growing. The company is developing artificial intelligence models for different types of robots, including robots that walk around and robot arms that stay rooted in one place.