In the second half of last year, Superb AI quietly launched an unusual real-world data collection effort that has since turn into a cornerstone of its push into physical AI. The South Korean startup rented 50 residential properties over approximately six months to capture everyday household activities for robot training. This initiative, described in company materials as gathering “robotic viewing” data, aimed to build a dataset reflecting authentic Korean home environments where future service robots might operate.
The effort resulted in 1.08 million video frames depicting 50 distinct domestic scenarios — such as setting tables, loading dishwashers, and navigating cluttered living spaces — each repeated two to three times to ensure robustness. These recordings were not made in labs or simulated settings but in actual rented homes across Korea, chosen to reflect regional variations in layout, lighting, and furniture arrangement. Superb AI’s goal was to train its vision foundation model, dubbed ZERO, to recognize and respond to real-world conditions without requiring extensive retraining for each new environment.
This approach aligns with a broader industry shift toward data curation over sheer volume. As Superb AI’s CTO Cha Moon-soo explained in a 2025 interview, “It’s not about how much you place in, but what you put in.” The company’s proprietary data selection process filters raw input to retain only the most informative examples — reducing 100 million raw frames to approximately 900,000 high-value samples. This curated approach enabled the development of ZERO using just eight A100 GPUs over eight months, achieving performance comparable to models trained on far larger computational budgets.
ZERO, introduced in June 2024 as an industry-specific vision foundation model, forms the technical core of Superb AI’s expansion into physical AI systems. Unlike conventional AI that relies on task-specific training, ZERO is designed to generalize across visual domains, allowing robots to interpret novel scenes — such as a spilled liquid or an obstructed pathway — using contextual understanding rather than predefined rules. The model has since been integrated into Superb AI’s video analytics platform, Superb VA, which analyzes live camera feeds to detect safety hazards like falls, fires, or unattended stoves.
In April 2026, Superb VA received official recognition as an innovative product from South Korea’s Public Procurement Service, enabling direct government contracts without competitive bidding for up to three years. The designation highlights the system’s ability to work with existing CCTV infrastructure, eliminating the need for costly hardware upgrades. According to CEO Kim Hyun-soo, this accreditation lowers adoption barriers for public institutions seeking to implement AI-driven monitoring for workplace safety and smart city initiatives.
The home-rental data collection effort underscores a growing realization in robotics: physical intelligence depends not just on algorithmic sophistication but on the quality and relevance of training data. By focusing on Korean residential settings — including apartment complexes common in urban centers like Seoul and Busan — Superb AI aims to address a gap in global AI datasets, which often underrepresent East Asian domestic environments. This localization strategy could improve the real-world applicability of service robots in aging societies where demand for home assistance is rising.
While the company has not disclosed the exact financial terms of the property leases or the total cost of the six-month filming campaign, the scale of the operation suggests a significant investment in ground-truth data acquisition. Such efforts are increasingly vital as robots transition from controlled industrial settings to unpredictable domestic and commercial spaces, where variability in lighting, occlusion, and human behavior poses persistent challenges to perception systems.
Looking ahead, Superb AI plans to leverage its ZERO model and proprietary data pipeline to expand into vertical-specific AI solutions for manufacturing, logistics, and healthcare. The startup’s participation in LG AI Research’s private foundation model consortium further signals its ambition to contribute to national AI infrastructure. For now, the 1.08 million frames gathered from those 50 rented homes remain a foundational asset — silent proof that teaching robots to understand our homes begins with watching us live in them.
Keep reading