LG Electronics Runs Robot Data Factory, Plans Expansion to 200 Robots

LG Electronics is operating the LG Data Factory to prepare robots for commercialization, with plans to expand its infrastructure from about 100 robots currently in operation to 200 by the end of this year.

The facility serves as a “robot training center,” supporting data generation, refinement, training and validation. Large volumes of training data are essential for robots to move like people, but collecting usable examples remains a demanding process.

Only about 30% of manually collected data is usable

According to Park Gi-beom of LG Electronics, teleoperation—collecting data while a person directly controls a robot—requires substantial time and money. Skilled operators must guide robots to create reference examples for training. Although the usable share varies by operator, only about 30% of the data collected this way can actually be used for training, he said.

Methods used to supplement the shortfall include generating simulation data in virtual environments, retargeting technology that converts human movements into robot movements, and augmentation techniques that artificially expand datasets.

LG Electronics is pursuing a strategy using Auto-Ops technology, which combines vision AI and imitation learning to let robots produce their own reference examples. The company aims to reduce human involvement and accelerate data generation through this approach.

A continuous cycle of training and validation

The LG Data Factory aims to establish a “Data Flywheel,” with data generation, refinement, training and validation forming a continuous cycle. When a robot deployed in the field encounters an unexpected environment, described as a domain gap, the cycle returns to data collection and further training.

The facility has IT infrastructure and networking capacity to process large volumes of data, extending its role beyond that of a research laboratory. It combines four types of Wi-Fi to handle the substantial bandwidth requirements of video data and has infrastructure for sharing large amounts of GPU resources.

Park said this supports a staged training system that progresses from simulation to validation in actual operating environments. He likened the simulation stage to a practice exam and real-world validation to the college entrance exam.

Data tailored to LG Electronics’ own robots

LG Electronics’ strategy centers on directly producing data optimized for its own robots’ physical configuration, or Form-factor, to improve model accuracy. The company says data reflecting the characteristics of its robots offers better training efficiency and accuracy than externally purchased data. It sees this as a foundation for enabling customers to train robots for specific tasks with less data and time.

A One LG team brings together capabilities from LG Electronics, an AI research institute, LG CNS and LG Innotek at the LG Data Factory. LG Innotek handles camera interfaces, while LG CNS delivers robots to commercial and logistics spaces and supports training tailored to customer requirements. The AI research institute supports the development of proprietary models.

Building on that collaboration, LG Electronics continues to cooperate with global technology companies, including NVIDIA and MS, on an ecosystem spanning hardware, software and cloud infrastructure.

Source: original video (YouTube)