The LG AI Research consortium, with Superb AI serving as a core partner, has passed the second evaluation of Korea’s Sovereign AI Foundation Model Project, led by the Ministry of Science and ICT (MSIT), and advanced to Phase 3. Within the consortium, Superb AI is the only non-LG affiliate directly involved in developing the high-performance AI foundation model, with responsibility for the vision, video, and Physical AI domains of K-EXAONE, the consortium’s large-scale foundation model.

This post looks at what the second evaluation was designed to test and how Superb AI has contributed to K-EXAONE through the work completed in Phases 1 and 2. For the official announcement, see the press release.

Beyond Benchmarks: Testing Whether the Model Works in Practice

The second evaluation went beyond benchmark performance and expert review. It also included a public evaluation in which 200 members of the public used the models directly and scored their experience. In other words, the evaluation assessed not only performance on predefined metrics, but also how useful the models are in practice. Three of the four finalist teams passed this stage and advanced to Phase 3.

The LG AI Research consortium has now cleared every stage of the program: selection as a finalist team in August 2025, first place across benchmark, expert, and user evaluations in the first assessment in January 2026, and now the second evaluation.

The consortium brings together 11 organizations, including LG AI Research, Superb AI, LG Uplus, LG CNS, FuriosaAI, FriendliAI, ESTsoft, ESTaid, Hancom, Wrtn Technologies, and Elice Group.

Progress of Korea’s Sovereign AI Foundation Model Project

The Eyes and Hands of Physical AI: Superb AI’s Role

Physical AI needs two things: eyes to perceive the world and data that teaches it how to act. Superb AI is building both to help K-EXAONE evolve beyond text into a model that can understand the physical world.

In Phase 2, Superb AI was responsible for advanced vision technologies, including visual understanding and 3D spatial interpretation. The team also built a multimodal data foundation by capturing and processing human actions from both first- and third-person perspectives and pairing them with natural-language labels. This gives the model a foundation for learning not only what an action looks like, but also the intent and context behind it.

Superb AI’s role across the vision, video, and Physical AI domains of K-EXAONE (conceptual illustration)

Phase 1: Building the Foundation for Korea-Specific Action Data

In Phase 1, Superb AI captured 1.08 million RGB-D frames across 50 Korean home environments and turned 300,000 of them into training- and validation-ready assets.

A multiview rig of 17 timecode-synchronized GoPro cameras captured both third-person views covering the full environment and first-person views designed to represent a robot’s perspective. The team staged approximately 7,500 demonstrations across 50 household-task scenarios, including cooking, dishwashing, and organizing.

The key challenge was selection. By simple calculation, the full volume of captured footage amounted to roughly 400 million frames. Using Auto-Curate, Superb AI removed repetitive segments with little training value to produce 1.08 million source frames, then narrowed those down again to 300,000 frames centered on edge cases with higher expected training value.

The videos were also captioned with the intent and causal context behind each action—for example, “grasped a glass with the right hand in order to drink water.” This allows the model to learn not only what action occurred, but why it occurred.

Phase 2: Turning Captured Data into Simulation-Ready Assets

Phase 2 was not about collecting more data. It was about turning the real-world data already captured into digital assets that robots can learn from in simulation.

Superb AI built three core types of digital assets: spaces, actions, and objects:

• Space assets: Superb AI reconstructed 50 Korean residential environments using 3D Gaussian Splatting at photorealistic fidelity. Physical properties such as floor friction and wall rigidity were then added so robots could use the environments for navigation and collision testing.
• Action assets: The team converted 5,000 household-task scenarios into SMPL-based motion data. Because body shape and pose are represented separately, a motion created once can be retargeted to digital humans and robots with different body types.
• Object assets: Superb AI converted 10,000 commonly handled objects into interactive digital assets. Moving components such as drawers and refrigerator doors were given hinge information and ranges of motion, allowing robots to practice opening and closing them in virtual environments.

When these three asset types come together, they enable a synthetic data pipeline that can generate thousands of variations from a single source by changing lighting, object placement, and movement paths.

Even rare scenarios that are difficult or unsafe to stage in the real world—such as cookware catching fire—can be generated safely as training data.

Phase 3 and What Comes Next

Superb AI will continue to be responsible for the vision, video, and Physical AI domains of K-EXAONE in Phase 3.

Building on the real-world data foundation established in Phase 1 and the digital asset pipeline created in Phase 2, Superb AI will continue working with the consortium to advance K-EXAONE toward a model capable of understanding and acting in the physical world.

Superb AI is a Vision Intelligence company that transforms visual data from industrial environments into actionable intelligence for enterprises.