August 2026 Summaries

South Korea’s expanding video analytics market faces a tension between demand for AI training data and strict privacy rules governing footage containing identifiable people, making de-identification a practical alternative to using original video. A vision AI company worked with Superb AI to develop pedestrian detection and later video tracking datasets from de-identified images, balancing privacy protection with the visual detail needed for accurate annotation. Superb AI established de-identification standards and labeling guidelines during project kickoff, incorporated the customer’s existing pre-labeling results, annotated pedestrians with bounding boxes and attributes, and converted completed labels into the format required for the customer’s training pipeline. The subsequent tracking project added consistent person IDs across video frames but could proceed quickly because the initial standards and team were already in place. The project produced privacy-compliant, documented training data while allowing the customer’s researchers to focus on model development rather than large-scale manual annotation.

Superb AI’s final series installment describes a proof-of-concept synthetic-data pipeline that combines reusable action, environment, and physics-ready object assets in a simulator to generate labeled training data for Physical AI. Compared with costly real-world recording, synthetic scenes can be rendered quickly and repeatedly, with domain randomization varying lighting, materials, viewpoints, and object placement while automatically producing labels such as bounding boxes, segmentation masks, depth, optical flow, pose, and 3D boxes. Validation currently relies on iterative visualization and error-specific tests to ensure digital humans interact physically correctly with floors and walls, though independently extracted spatial and motion data still require post-processing alignment. The work emphasizes that photorealism alone is not the goal; useful synthetic data must transfer effectively to real settings by accounting for physical behavior, sensor conditions, and rare edge cases. The company positions its reusable asset model as a process-efficient complement to large-scale physical data collection, with full production planned for the next phase.

A public-sector institution referred to as Institution B used a three-month Information Strategy Planning consulting engagement with Superb AI to prepare for adopting generative AI in a network-isolated environment handling sensitive data. The work involved interviews, surveys, and a data-center assessment to identify workflow needs, security classifications, current infrastructure, and implementation constraints. Candidate commercial and custom large language models were evaluated using shared criteria such as parameter size, context length, training data, and response quality, while the target architecture separated processing paths for general and higher-security data. The proposed on-premises design combined LLMs with retrieval-augmented generation for using internal documents and considered integration methods such as the Model Context Protocol. The engagement also defined needed software components, functional requirements, data-management practices, staffing and schedule estimates, and procurement reference materials, emphasizing that organizational diagnosis and system design should precede final model selection and full implementation.

Company R, an autonomous robot service provider, has worked with Superb AI since 2022 to build training-data pipelines for outdoor delivery robots that navigate complex urban environments using LiDAR and multi-camera systems. The collaboration addresses challenges such as aligning 3D point clouds with camera footage, tracking objects across sequential frames, labeling pedestrians, vehicles, traffic lights, and traversable space, and accommodating distorted fisheye imagery. Superb AI’s platform manages data intake, labeling, quality review, batch delivery, custom export formats, and evolving specifications, while supporting both existing customer data and projects involving on-site collection partners. The partnership also operates under personal-data consignment and annual security-review requirements, including access controls, training, retention-based deletion, and subprocessor oversight. The case presents sustained, adaptable data operations as important infrastructure for improving perception AI as robot service areas, scenarios, and data requirements expand.

As part of Korea’s Dokpamo Sovereign AI Foundation Model Project, Phase 2 converted imagery from 300,000 real Korean home-environment frames into 10,000 interactive household-object assets intended to train robots for manipulation tasks. The assets include pixel-level segmentation and, for eight articulated categories such as cabinets, refrigerators, drawers, wardrobes, and curtains, mechanical properties including motion axes and ranges; rigid objects such as cups and plates were modeled separately. An AI-assisted workflow used SAM and computer-vision methods for initial labeling, a vision-language model for review, and human reviewers for final quality control, applying a dataset-wide passing threshold of at least 0.7 IoU and 90% semantic accuracy. Occlusion in realistic household scenes was identified as the principal segmentation challenge, while curtains required a simplified sliding-motion model to balance realism with computational cost. The project positions its focus on Korean home contexts and articulated objects alongside related efforts such as Meta’s SAM 3D and AgiBot World, with a future phase planned to combine spatial, object, and behavior assets into a synthetic-data pipeline.

A manufacturer of industrial automation components adopted Superb Platform to build AI datasets for PCB component identification and defect detection after open-source labeling tools became difficult to manage at production scale. Its small internal research team needed to combine bounding boxes, polygons, and rotated bounding boxes for varied component shapes and defects while keeping sensitive product data in-house. The platform enabled migration of existing annotations, centralized class and project management, automated the identification of potentially incorrect labels through Auto-Curate, and supported object-level model diagnosis, dataset slicing, and similar-image search. By integrating labeling, quality review, dataset exploration, and progress tracking, the workflow reduced the operational burden on team members who handled both annotation and management, with the case emphasizing that dataset quality and labeling operations are central to manufacturing AI performance.

Hyun Kim describes how Phase 2 of Korea’s Sovereign AI Foundation Model Project reconstructed 50 Korean homes into simulator-ready digital environments for robot learning, rather than producing conventional panoramic or viewing-focused scans. Using 3D Gaussian Splatting, the team created freely navigable visual backgrounds from ordinary camera footage while prioritizing accurate real-world scale, ground-plane calibration, and structural reliability over visual fidelity. Since 3DGS does not separate scene objects, interactive items such as doors, drawers, dishes, and clothing are added as distinct physical meshes that support collision and manipulation, while wall and floor planes help reduce physically implausible interactions. The work exposed reconstruction difficulties in narrow rooms, and a related office digital-twin project led to improvements in multi-person and robot avoidance behavior in NVIDIA Isaac Sim. The effort is presented within a growing global competition to build physical AI training infrastructure, where reusable, simulation-compatible digital assets may determine how effectively real-world facilities can support robot training.

Superb AI has released ZERO 2.2, an updated industry-focused Vision Foundation Model available through the Labson Superb Platform and AWS Marketplace, designed to detect previously unseen objects using text prompts or boxed image examples without collecting data, labeling images, or retraining models. The update improves text-based detection accuracy, reduces missed detections, supports Korean and Japanese prompts, recognizes descriptive phrases such as color and clothing attributes, and provides clearer confidence-score separation to simplify threshold selection in operational settings. Example-image prompting enables users to identify visually similar or unnamed products by selecting one instance, supporting applications in manufacturing inspection, logistics, retail analytics, and security monitoring. Superb AI reports that the model was trained on roughly 900,000 curated industrial images and outperformed open models across 37 industrial datasets, while its underlying approach won first place in CVPR 2026’s Foundational Few-Shot Object Detection Challenge with an mAP of 53.9. Users can test the model without installation on the Labson platform or deploy it as a usage-based Amazon SageMaker endpoint through AWS Marketplace.

A global electronics and precision equipment manufacturer is working with Superb AI to reduce visual-inspection bottlenecks at overseas plants without replacing existing automated optical inspection hardware. Rule-based AOI and AFVI systems struggle with irregular surface defects such as scratches, dents, contamination, and delamination, causing either excessive false positives that require manual review or missed defects that may reach customers. The company is testing a deep learning system that performs secondary review of equipment-generated defect candidates, initially focusing on the most frequent and quality-critical defect categories identified through analysis of more than 100 defect codes. Using Superb AI’s platform, the project applies iterative data labeling, model training, and evaluation, including a Transformer-based model for fine-pattern recognition. Its confidence-based workflow automatically passes high-confidence cases, sends medium-confidence cases to human experts, and uses uncertain cases and human review results for retraining, with the aim of gradually increasing straight-through inspection rates, reducing dependence on manual labor, and creating reusable inspection data for quality improvement.

Superb AI describes its Phase 2 work in Korea’s Sovereign AI Foundation Model Project, in which it converted 7,500 recordings of household activities across 50 Korean homes into 5,000 quality-approved 4D human behavior assets for robot learning. The assets use SMPL fitting to represent body shape and pose separately over time, allowing captured motions to be retargeted to digital humans or robots with different body types. The company contrasts human behavior observation with robot teleoperation, arguing that observation data offers scalable diversity, whole-body coordination, tool-use context, and intent for pretraining, while teleoperation provides robot-coordinate precision for fine-tuning; simulation and retargeting connect both methods. Challenges such as occlusion, limited camera coverage, and constrained rooms were addressed through anatomically informed pose estimation, temporal smoothing, interpolation, and discarding low-confidence segments, resulting in successful processing of more than 98% of recordings. Each behavior can be augmented through multiple body models, viewpoints, and backgrounds, and captions generated with vision-language models and reviewed by experts add a language layer intended to support multimodal learning.

Korea Land & Housing Corporation (LH), the country’s largest public housing provider, partnered with Superb AI to automate residential defect intake and classification using a Vision-Language Model that interprets photos and complaint text together. Facing roughly 300,000 defect images monthly, inconsistent classifications driven by varying agent experience, and a five-dimensional taxonomy producing more than 18,000 possible categories, LH adopted a single-model system that classifies space, materials, defect type, construction trade, and work type in real time. The system identifies issues such as cracks, leaks, mold, damage, and staining, provides natural-language explanations for its classifications, and was designed to meet targets of at least 90% detection accuracy, under 10% false positives and negatives, and processing in less than one second per image. Beyond reducing reliance on manual review and improving repair routing, the initiative creates structured defect data that LH can use to identify recurring problems, plan maintenance and budgets, and move toward preventive asset management, while offering a potential model for other public services handling large volumes of unstructured image-based reports.

Korea's Sovereign AI Foundation Model Project, known as Dokpamo, aims to create 3D spaces, 4D behaviors, and intelligent objects as foundational components for virtual environments from which robots can learn. This initiative, led by the Korean government, addresses the challenge of training robots in local environments, as data sourced from elsewhere may not be effective. A key element of the project is the Physical AI Data Factory, which converts real-world data into digital assets for simulators, generating large volumes of labeled training data. The project underscores the importance of acquiring high-quality real-world data to produce synthetic data that accurately reflects reality. As Korea emphasizes the development of data infrastructure for physical AI, similar initiatives are underway globally, including in China and Germany, highlighting a competitive race to gather physical-world learning data. Superb AI is actively contributing to this effort in Korea and encourages collaboration through their enterprise-level training data platform, which streamlines the preparation of high-quality training datasets.