"AgiBot speeds up the learning process by having a human worker guide the robot through a task, which provides a foundation for it to then learn by itself. Before cofounding AgiBot, chief scientist Jianlan Luo did cutting-edge research at UC Berkeley, including a project that involved robots acquiring skills through reinforcement learning with a human in the loop. That system was shown doing tasks including placing components on a motherboard."
"Feng says that AgiBot's learning software, called Real-World Reinforcement Learning, only needs about ten minutes to train a robot to do a new task. Rapid learning is important because production lines often change from one week to the next, or even during the same production run, and robots that can master a new step quickly can adapt alongside human workers."
Reinforcement learning for robots that require improvisation demands large amounts of training data and cannot be perfected solely in simulation. AgiBot accelerates learning by having human workers guide robots through tasks to provide a foundation for autonomous learning. Chief scientist Jianlan Luo previously led related research at UC Berkeley involving human-in-the-loop reinforcement learning for tasks such as placing motherboard components. AgiBot's Real-World Reinforcement Learning trains robots in about ten minutes, enabling rapid adaptation to production line changes. The company operates a robotic learning center that pays teleoperators to control robots and generate training data. AgiBot develops models for humanoids and stationary robot arms within China.
Read at WIRED
Unable to calculate read time
Collection
[
|
...
]