"Only 30 out of 100 Survive"... LG Electronics Operates 'Data Factory' for Robot Learning

In order for robots to move like humans, a vast amount of training data is essential. However, the data robots encounter in the field is like unrefined ore, and the process of turning it into an "answer sheet" that can be used for actual training requires enormous cost and time. To solve this problem, LG Electronics has built "LG Data Factory," a kind of robot training center that generates, refines, trains, and even verifies data.

"Only 30 out of 100 can be used"... The high threshold of robot data

The data factory operated by LG Electronics is like an "academy" for robots. Park Gi-beom, a person in charge at LG Electronics, defined the data factory as "a robot training center and a place that creates study materials by generating, refining, training, and verifying data." Currently, about 100 robots are deployed here, and the scale is expected to expand to 200 robots by the end of the year. Park stated, "We started in the second half of the year, and the speed is so fast that 200 units will be ready soon," adding, "The actual mass production scale will be much larger than this."

The biggest challenge in robot learning is the quality of data. According to Park, the 'Teleoperation' method, where a person directly controls a robot to collect data, provides the highest quality data, but because it requires the labor of skilled workers, the time and cost are very high. In fact, even if 100 pieces of data are collected through teleoperation, depending on the operator, the effective data that can actually be used for learning is only about 30%.

To overcome this limitation, LG Electronics secures data through three paths. First is the data obtained by directly controlling the robot; second is data using 'Retargeting' technology, which converts human movement into robot movement; and third is the 'Synthetic Data' method, which amplifies data in a virtual environment. Retargeting is the task of mapping human video to robot actuator values. In particular, recently, they are focusing on increasing the speed of data generation and reducing human intervention through 'Auto-ops' technology, which combines vision AI and imitation learning to allow robots to move on their own and create answer sheets. However, Park added, "While Auto-ops is fast and can accumulate data without humans, it has the disadvantage of lacking data diversity."

Data Flywheel and 'Auto-ops'... Technology to reduce the field gap

Solving the 'Domain Gap' that occurs when robots apply what they have learned to actual fields is also a core task of the data factory. If unexpected variables occur, such as different lighting in a factory or slight differences in the height and number of shelves, the robot may malfunction. To address this, LG Electronics operates a 'Data Flywheel' structure that generates data, trains it, sends it to the field, and then collects data again to re-train.

Park compared this process to the CSAT (College Scholastic Ability Test). "If the training and mid-training processes are several mock exams, the field Proof of Concept (PoC) is like the CSAT," he explained. "Through mock exams, various variables are learned, and in the final stage, completeness is increased through post-training tailored to the actual field." The amount of data required for learning varies greatly depending on the type of task; for example, tasks such as moving a cup can be effective with about 100 hours of data based on post-training.

Instead of purchasing data from outside, LG Electronics adheres to a strategy of building data using its own machines. This is because while external data may be cheap, its accuracy drops if it does not match the characteristics (Form Factor) of the robot machines directly made by LG Electronics. Robot learning centers on two sets: 'what task will be performed' and 'what state my body (machine) must be in at that time,' and data made with their own machines has much higher accuracy. Park emphasized, "If we make a model specialized for our machines, it is possible to perform sophisticated learning with much less data and time when customers perform specific tasks in the field."

Securing leadership in Physical AI through 'One LG' synergy

The data factory is a space where the capabilities of LG Group are concentrated. The independent model development and mid-training expertise of AI researchers, the high-performance camera technology (video signal processing) of LG Innotek, and the customer-customized system construction and delivery capabilities of LG CNS have all come together. This 'One LG' system serves as the foundation to quickly supply completed robots by integrating hardware and software.

Recently, officials from NVIDIA visited the data factory and expressed surprise at the speed of LG's preparation, even using the expression "Incredible." Park explained, "Unlike American companies that rent Airbnb to secure space for robot testing, LG has simultaneously built infrastructure capable of handling large-scale data and an environment capable of solving mass production stage problems." He also views LG as a partner that can provide to NVIDIA as an E2E (End-to-End) assembled entity, and as a cooperation model that can provide cloud infrastructure for robot operation to MS.

Ultimately, the deciding factor in the era of Physical AI depends on 'who can more efficiently secure higher quality data.' Based on the technological prowess built through the data factory, LG Electronics plans to realize the 'Zero Labor Home' vision that replaces domestic labor.

Source: original video (YouTube)