具身智能机器人技术

摘要：提出了一种用于智能制造的具身智能机器人技术。提出的具身智能机器人通过主动感知环境、自主学习和自主决策来执行拟人化任务，其“大脑”是以强化学习为核心的智能体，感知部分具有双目视觉、空间六维力感知和本体状态感知等多模态感知能力；通过“感知-行动”反馈环，构建了一个控制周期为1 ms的机器人实时控制系统。在中兴通讯的5G智能制造工厂中部署了具身智能机器人进行实际生产，用来取代工人插拔5G小站产品的RJ45插头和光模块。实践表明，具身智能可打通全自动化生产线的最后断点，是一项在智能制造中有广阔应用前景的机器人技术。

关键词：具身智能；强化学习；机器人；多模态感知

Abstract: An embodied intelligent robot technology in industrial intelligent manufacturing is proposed. The robots can learn the policy, make decisions, and act autonomously to complete anthropomorphic tasks through active perception and environmental interaction. The embodied intelligent robot uses a reinforcement learning agent as its “brain” with a real-time control system of 1 ms control cycle with a multimodal "perception-action" feedback loop, which includes stereo vision, spatial six-dimensional force sensor, and robot proprioception. The system was subsequently deployed in our 5G intelligent manufacturing factory to replace workers, to plug and unplug network RJ45 crystal heads and optical modules for 5G small station manufacture, which eliminated our last breakpoint of automation. The practice shows that embodied intelligence is a promising robot technology in the future manufacturing industry.

Keywords: embodied AI; reinforcement learning; robotic; multimodal perception