LG电子扩大与英伟达的机器人数据合作
LG电子正在扩大与英伟达的合作,以构建机器人训练数据。在首尔良才的数据工厂,该公司计划将真实世界任务数据与虚拟数据相结合。
原文深度解读
LG电子正在扩大与英伟达的合作,以构建机器人训练数据。在首尔良才的数据工厂,该公司计划将真实世界任务数据与虚拟数据相结合,到年底确保10万小时的训练数据。文章涉及LG电子、英伟达、LG CNS、LG Innotek等主体,时间点为2025年8月18日及前后。关键数字包括10万小时训练数据、10000平方米工厂面积、年底前部署数百台机器人。这与电子元器件供应链相关,因为机器人训练数据涉及传感器、执行器等元器件的应用,且LG电子计划开发关键组件如执行器。
- LG电子于8月18日邀请英伟达高管访问首尔良才数据工厂,讨论机器人合作,此前四天双方在美国英伟达总部签署了关于机器人、AI工厂和移动出行的谅解备忘录。
- 良才数据工厂目标年底全面运营,LG电子部署自研LG CLOiD机器人进行数据生成、收集和训练,设施包括模拟家庭、制造和物流场景。
- 数据通过英伟达机器人解决方案(包括Omniverse库、Cosmos开放世界模型和Isaac开放机器人开发平台)进行增强和合成。
- LG电子计划利用制造和物流现场积累的任务数据及熟练工人的技能作为训练数据,构建数据飞轮,机器人反复训练真实和合成数据,性能提升反馈到进一步训练。
- 良才数据工厂从地下一层到地上三层共四层,总建筑面积10000平方米,年底前计划将部署机器人数量增加到数百台。
- 结合真实和虚拟数据的训练数据预计年底达到10万小时,相当于约12年,LG电子计划用于推进人形机器人的机器人基础模型(RFM)。
供应链影响
- LG电子扩大机器人业务可能增加对执行器等关键组件的需求,其自研和生产能力可能影响相关元器件供应链。
- 数据工厂部署数百台机器人可能带动传感器、计算芯片等电子元器件的采购,具体数量取决于机器人配置。
- 与英伟达的合作可能促进AI相关硬件(如GPU)在机器人训练中的应用,进而影响高性能计算元器件的供应链。
- LG电子整合机器人业务结构可能改变其采购策略,对现有供应商的依赖可能变化。
- 机器人训练数据的积累可能提升机器人性能,进而影响未来机器人产品的市场推广,间接影响元器件需求,但具体影响取决于产品化进程。
系统解读,仅作为判断线索,不构成备货、出货、涨价或投资建议。
英伟达高级总监Madison Huang于8月18日访问LG电子首尔数据工厂,并在LG CLOiD上留下签名留言“AMAZNIG LG”。(照片:LG电子)
LG电子正在扩大与英伟达的合作,以构建机器人训练数据。在首尔良才的数据工厂,该公司计划将真实世界任务数据与虚拟数据相结合,到年底确保10万小时的训练数据。
LG电子于8月18日表示,已邀请英伟达高管访问其良才数据工厂,讨论机器人合作。此次会面是在双方于英伟达美国总部签署谅解备忘录(MOU)四天后进行的,该备忘录涵盖机器人、人工智能(AI)工厂和移动出行方面的合作。
双方考察了良才数据工厂的运营情况,该工厂正以年底全面运营为目标进行建设。LG电子正在部署其内部开发的LG CLOiD机器人,用于生成、收集和训练数据。该设施包括模拟家庭以及制造和物流场所的空间。LG CLOiD在模拟家庭环境中执行清洁任务。在仿照LG电子田纳西州洗衣机工厂的区域,机器人执行包括运输、装载和组装组件在内的任务。
机器人也已部署在具有LG CNS物流自动化解决方案和LG Innotek机器人手训练环境的区域。通过这些操作收集的数据使用英伟达机器人解决方案进行增强和合成,包括Nvidia Omniverse库、Cosmos开放世界模型和Isaac开放机器人开发平台。
LG电子计划使用在制造和物流现场积累的任务数据以及熟练工人的技能作为训练数据。该公司正在构建一个数据飞轮,机器人在其中反复训练真实和合成数据,并将由此产生的性能改进反馈到进一步训练中。
良才数据工厂将从地下一层到地上三层共四层,总建筑面积10000平方米。LG电子计划到年底将该设施部署的机器人数量增加到数百台。
结合收集的真实数据与增强和合成的虚拟数据的训练数据预计到年底达到10万小时,相当于约12年。LG电子计划使用这些数据来推进其人形机器人机器人基础模型(RFM)。
LG电子还在加强其机器人组织。该公司上个月成立了一个直接向首席执行官(CEO)汇报的机器人业务中心,以监督其机器人运营。它正在将产品线从工业和商业机器人扩展到家庭机器人,同时建设开发和制造关键组件(如执行器)的能力。该公司还在建立一个涵盖数据工厂、硬件和软件的集成机器人业务结构。
“我们将通过‘One LG’协同效应确保物理AI的竞争力,该协同效应汇集了集团的核心能力以及与全球合作伙伴的战略合作,”LG电子CEO Ryu Jae-cheol表示。
Madison Huang, a senior director at Nvidia, visits LG Electronics' Data Factory in Seoul on Aug. 18 and leaves a signed message, "AMAZNIG LG," on LG CLOiD. (Photo: LG Electronics)
LG Electronics is expanding its collaboration with Nvidia to build robotics training data. At its Data Factory in Yangjae, Seoul, the company plans to combine real-world task data with virtual data to secure 100,000 hours of training data by the end of the year.
LG Electronics said on Aug. 18 that it invited senior Nvidia executives to its Yangjae Data Factory to discuss cooperation in robotics. The meeting came four days after the companies signed a memorandum of understanding (MOU) at Nvidia's US headquarters covering collaboration in robotics, artificial intelligence (AI) factories and mobility.
The companies reviewed operations at the Yangjae Data Factory, which is being built with the goal of reaching full operation by year-end. LG Electronics is deploying its internally developed LG CLOiD robot to generate, collect and train on data. The facility includes spaces replicating homes as well as manufacturing and logistics sites. LG CLOiD performs cleaning tasks in the simulated home environment. In an area modeled after LG Electronics' washing-machine factory in Tennessee, the robot performs tasks including transporting, loading and assembling components.
Robots have also been deployed in areas featuring LG CNS' logistics automation solutions and LG Innotek's robot-hand training environment. Data collected through these operations is augmented and synthesized using Nvidia robotics solutions, including Nvidia Omniverse libraries, Cosmos open-world models and the Isaac open robotics development platform.
LG Electronics plans to use task data accumulated at manufacturing and logistics sites, as well as the know-how of skilled workers, as training data. The company is building a data flywheel in which robots repeatedly train on real and synthetic data, with the resulting performance improvements fed back into further training.
The Yangjae Data Factory will span four floors from one basement level to three floors above ground, with a total floor area of 10,000 square meters. LG Electronics plans to increase the number of robots deployed at the facility to several hundred by the end of the year.
Training data combining collected real-world data with augmented and synthesized virtual data is expected to reach 100,000 hours by year-end, equivalent to about 12 years. LG Electronics plans to use the data to advance its robot foundation model (RFM) for humanoid robots.
LG Electronics is also strengthening its robotics organization. The company established a Robotics Business Center reporting directly to the chief executive officer (CEO) last month to oversee its robotics operations. It is expanding its product lineup from industrial and commercial robots to home robots while building capabilities to develop and manufacture key components such as actuators. The company is also establishing an integrated robotics business structure spanning the Data Factory, hardware and software.
"We will secure competitiveness in physical AI through 'One LG' synergies that bring together the group's core capabilities and strategic collaboration with global partners," LG Electronics CEO Ryu Jae-cheol said.