LG Electronics uses Nvidia AI to automate washing machine assembly

Automating heavy manufacturing

LG Electronics is currently testing a humanoid robot to perform specific tasks on a washing machine production line in the United States. The robot is tasked with hanging washing machine top plates onto coating hangers, a job that requires significant physical effort. By automating this specific motion, the company hopes to reduce the musculoskeletal strain and fatigue experienced by human workers who perform the task repeatedly.

To prepare the robot for this role, LG spent about a year at its Tennessee plant validating the movements. The company analyzed the motions of skilled human workers to create a baseline for the robot to follow. They then used a combination of remote operation and simulations to gather the necessary training data for the AI model.

The project represents a significant step in gauging the commercial viability of physical AI in the home appliance industry. While the robot is currently focused on a single task, the goal is to create a foundation for broader automation. LG intends to use the insights gained here to expand robot deployment to other production bases and different manufacturing processes in the future.

Bridging the virtual and real

A major challenge in robotics is the 'sim-to-real' gap, where a robot performs well in a computer simulation but fails in a real factory. To solve this, LG utilized Nvidia's Isaac GR00T robot AI model and the Cosmos video generation model. Cosmos helps create diverse training backgrounds to ensure the robot does not struggle with changes in lighting or surrounding conditions.

The company also employed Isaac Sim to build a virtual production process and used TensorRT to speed up the robot's decision-making time. When the robot fails a task on the actual factory floor, the engineers reproduce that specific failure in a virtual space. They then supplement the data and retrain the robot until it can successfully navigate the situation.

Park Jung-seop, head of LG Electronics' Robotics Intelligence Team, noted that securing performance under changing production plans is a primary task. The robot must be able to repeat the same motion even if parts shift slightly or the environment changes. This iterative process of failing, simulating, and retraining creates a virtuous cycle of data to improve reliability.

The LG Data Factory

LG Electronics is establishing a centralized system called the 'LG Data Factory' to manage its robotics development. This system integrates every step of the process, including data collection, generation, model training, on-site validation, and final deployment. By centralizing these functions, the company can more efficiently scale its AI capabilities across different products.

The company has set an ambitious goal to secure 500,000 cumulative hours of training data by the end of this year. This massive dataset will serve as the backbone for their robotics initiatives. Having a large repository of diverse data allows the company to train robots for a wide variety of tasks beyond just washing machine assembly.

This infrastructure also allows for faster deployment of new robots, such as the wheeled LG CLOiD which was recently deployed at the Tennessee plant. By building a reusable framework, LG can move from validating a single task to implementing complex, multi-step manufacturing workflows. The data factory ensures that every success and failure on the floor contributes to the overall intelligence of the fleet.

Next steps for bipedal robots

While the current project focuses on a humanoid robot for assembly, LG plans to unveil a next-generation bipedal humanoid robot in the first quarter of next year. Unlike the wheeled CLOiD, a bipedal robot can navigate stairs and narrow passages. This mobility allows it to reach areas of a factory that are currently inaccessible to wheeled machines, broadening its potential applications.

However, several hurdles remain before these bipedal robots can reach mass production. Engineers must still solve issues regarding walking stability, battery life, and strict compliance with safety regulations. The transition from a controlled production line to a dynamic factory floor requires the robot to be both agile and extremely safe around human coworkers.

For Nvidia, this partnership provides a large-scale case of embedding its simulation tools in an actual manufacturing site. This success in South Korea may lead to more references for other Asian manufacturing customers. As the technology matures, the focus will shift from simple repetitive tasks to complex coordination between multiple robots and human workers.