The real question is how effectively AgiBot’s algorithms can teach its robots fresh tricks. Using reinforcement learning to teach robots tasks that require improvisation typically requires a lot of training data, and research shows that it cannot be completely perfected in simulation.
AgiBot accelerates the learning process because the employee guides the robot through the task, which provides the basis for subsequent independent learning. Before co-founding AgiBot, chief scientist Jianlan Luo conducted cutting-edge research at the University of California, Berkeley, including design this involved robots acquiring skills through reinforcement learning with a human in the loop. This system is shown to perform tasks including placing components on the motherboard.
Feng says AgiBot’s educational software, called Real-World Reinforcement Learning, only takes about ten minutes to train the robot to perform a fresh task. Learning quickly is vital because production lines often change from week to week and even within the same production run, and robots that can quickly master a fresh step can adapt along with workers.
Training robots in this way requires a lot of human effort. AgiBot has robot learning center where people are paid to remotely control robots to lend a hand AI models learn fresh skills. The demand for this type of robot training data is growing, for some U.S. companies paying workers in places like India to perform manual work that serves as training data.
Jeff Schneider, a roboticist at Carnegie Mellon University who works on reinforcement learning, says AgiBot uses cutting-edge techniques and should be able to automate tasks with high reliability. Schneider adds that other robotics companies are likely experimenting with using reinforcement learning in manufacturing tasks.
AgiBot is something of a rising star in China, where interest in combining artificial intelligence and robotics is growing. The company develops artificial intelligence models for various types of robots, including walking humanoids and robotic arms that remain rooted in one place.