September 2026 Summaries

Selecting an AI CCTV vendor requires evaluating compatibility with existing cameras and NVRs, support for on-premises or edge deployment, independently measured false-alarm performance, definitions and flexibility of detectable events, retention of event and response records, maintenance terms, and proven results at comparable sites. The source emphasizes that demonstrations may not reflect real-world conditions, since lighting, network capacity, camera placement, and site layouts can affect performance after deployment. Organizations should assess whether current camera resolution and angles are adequate, determine server needs for non-cloud deployments, pilot systems to establish acceptable alert rates, and clarify whether new detection requirements require costly model retraining. It also recommends comparing vendors using consistent questions, validating claims at the intended site, and separating quotations into hardware, software, deployment, and operational costs to identify total ownership costs and potential savings.

Superb AI’s 93-second Physical AI brand film, “Right now, somewhere,” presents virtual replicas of Korean homes as safe training environments where robots can repeatedly practice tasks using 3D spaces, 4D action data, and interactive object assets created during Phase 2 of Korea’s Sovereign AI Foundation Model project. The film emphasizes that training environments must behave realistically as well as look realistic, showing that 5,000 of 7,500 recorded actions passed standards involving scale, ground alignment, and eight joint types, while rejected actions were excluded from training assets. It depicts robot failures as a productive feature of virtual practice rather than evidence of flawed data, using a conceptual counter to illustrate repeated training. Available in Korean, English, and Japanese without voice-over, the film continues the visual language of an earlier installment through on-screen text, counters, and graphics. Superb AI states that the underlying project produced 50 3D spaces, 5,000 qualifying 4D actions, and 10,000 object images, using 3D Gaussian Splatting backgrounds and physics meshes while excluding household-identifying information.

Industrial facilities in energy, shipbuilding, steel, and chemicals face common safety-monitoring challenges, including large hazardous areas, CCTV blind spots, fragmented systems for worker behavior and equipment anomalies, and the presence of rotating contractors. South Korea recorded 282 chemical accidents in 2025, with incidents rising during the summer months as heat can increase chemical volatility and storage-container pressure, underscoring the need for broader monitoring. The material describes sector-specific risks such as confined-space and height work in shipyards, outdoor blind spots at energy sites, heavy equipment near workers in steel plants, and fires, leaks, and PPE compliance in chemical facilities. It presents integrated AI video analytics as a way to use existing cameras to identify both behavioral and equipment-related events through a single interface, while noting that conventional systems may require months of training data and retraining for new hazards. Superb AI’s system is described as using a vision foundation model to add event types without task-specific retraining, support natural-language video search, operate on-premises for secure facilities, and be introduced through infrastructure assessment, event definition, pilot testing, and staged expansion.

Outsourced data-labeling prices vary widely because costs depend on annotation complexity, class counts and attributes, accuracy targets, quality-review stages, data conditions, security or de-identification needs, volume, delivery deadlines, and whether work is a one-time project or recurring operation. More detailed tasks such as polygons, segmentation, keypoints, and 3D cuboids generally require substantially more effort than image classification or bounding boxes, while specialized domains may require expert reviewers. Data curation, including removal of duplicate video frames, can reduce labeling volume and cost, although automated pre-labeling still requires human review for quality control and exceptions. To compare vendor quotes effectively, organizations should clarify their data-arrival pattern, required expertise, security constraints, review process, anticipated rework, and delivery terms, often beginning with a pilot batch before scaling production. The post also describes platform-based workflows and managed labeling services as options for organizations seeking to handle ongoing, high-volume visual-data annotation.

Vision AI data-pipeline total cost of ownership varies among in-house development, outsourcing, and platforms, with in-house costs dominated by staffing, outsourcing by per-item charges, and platforms by subscription fees and the volume of data requiring human review. A three-year planning horizon is recommended because models and datasets commonly require multiple updates, and organizations should compare itemized costs for upfront setup, staffing, data updates, and quality management rather than relying on headline pricing. Outsourcing may suit one-time dataset projects, while repeatable pipelines become more economical when products, cameras, data volumes, classes, or decision criteria change frequently. Prompt-based foundation models, automated labeling, and synthetic data can reduce initial manual-labeling needs, although they shift costs toward review, validation, simulator development, or platform operations. Organizations should estimate anticipated data updates, assess internal labeling and quality-review capacity, and determine whether tasks can begin without labeled data before selecting an approach; hybrid models can combine a platform for workflow governance with external labeling services for production-scale work.

South Korea is expanding defense AI infrastructure through GPU investments and plans for an integrated data center, but effective model development remains constrained by security rules, limited access to rare or sensitive targets, and the high cost of collecting and labeling real-world imagery. Synthetic data can supplement these gaps through compositing objects into real images, generative augmentation of conditions such as weather and lighting, and simulator-based generation of labeled 3D scenes. Superb AI describes using these approaches for defense and security datasets, including a simulator pipeline built from human-action samples, indoor environments, and object assets that can produce extensive variations. The approach is most applicable to rare events, unphotographable targets, and costly large-scale projects, while tasks with sufficient real-world data may not require it. Synthetic data should be curated carefully, combined with real data for training, and evaluated against real-world datasets to address the Sim-to-Real gap, with closed-network, on-premises deployment presented as a way to meet defense security requirements.

Superb AI describes three complementary methods for building robot behavior datasets: generating demonstrations in NVIDIA Isaac Sim, collecting teleoperated demonstrations with physical robots, and allowing a quadruped robot to gather camera data autonomously while patrolling. Simulation can produce large, varied datasets by changing environments, tasks, and objects, addressing the limited generalization of an earlier GR00T model trained on 1,000 episodes involving only a red cube in a sparse scene. Physical-robot collection has yielded 2,500 episodes combining synchronized head and wrist camera footage, robot joint states, and language instructions, with quality controls that discard incomplete or poorly visible recordings. The company preserves differences between commanded and actual joint states because the robot’s safety-limited low-level controller creates those differences during both training and inference. In the autonomous patrol demonstration, locomotion is learned through reinforcement learning, while the arm’s bottle-and-cup manipulation is computed with cuRobo motion planning rather than learned. Superb AI emphasizes that combining simulated and real-world data requires aligning their distributions, including by applying comparable physical constraints such as joint-velocity limits in simulation, and plans to expand its robot behavior data alongside its existing industrial visual-data collection.

Superb AI researchers demonstrated a method for deriving both camera movement and 3D hand motion from real first-person GoPro footage, aiming to create spatial action data for robot and Physical AI training without markers, studio setups, or simulated imagery. Using 1920×1080, 60 FPS video, the system combines RGB and GoPro IMU data for camera odometry, while separately using RGB data to reconstruct hand movements; the results are integrated into a Gaussian Splatting-based 3D reconstruction of the workspace. This shared coordinate system records not only how hands moved but also where actions occurred in the environment, with validation performed by comparing renderings from estimated camera poses against original footage. The camera-trajectory pipeline achieved a median position error of about 1.4 cm, rising to roughly 3 cm in more difficult segments, against an offline pseudo-ground-truth reference built with additional external-view imagery. Running at 27 FPS, the odometry component meets the stated threshold for real-time video inference and can also independently estimate a wearer’s six-degree-of-freedom path through an environment.

AI video analytics implementation has no standard price because costs vary with existing camera infrastructure, the number of video channels, required detection tasks, deployment model, and ongoing support needs. Major expenses include cameras and cabling when existing CCTV cannot be reused, GPU servers for video processing, software licenses, deployment, custom model development, and maintenance such as false-positive tuning and system expansion. Cloud deployments may lower upfront spending, while on-premises systems can be more economical over time and may be necessary where video cannot leave a site. General person and vehicle detection can often use existing models, but specialized events may require custom training; zero-shot models such as Superb AI’s ZERO aim to reduce initial data-labeling requirements and allow pilots with relatively few images. The recommended approach is to assess reusable infrastructure, define detection needs, test a limited pilot, then scale a validated configuration while evaluating total cost of ownership rather than only initial pricing. Companies in Korea may also be eligible for government support programs, including 2025 AI Voucher and Smart Factory initiatives offering up to KRW 200 million under specified conditions.

Superb AI, the sole non-LG affiliate directly developing K-EXAONE within the 11-member LG AI Research consortium, helped the group advance to Phase 3 of South Korea’s Sovereign AI Foundation Model Project after passing an evaluation that combined benchmarks, expert review, and feedback from 200 public users. The company leads the model’s vision, video, and Physical AI work, aiming to extend its capabilities from text-based understanding to perception and action in real-world environments. During Phase 1, Superb AI recorded household activities across 50 Korean homes using synchronized first- and third-person cameras, then used its Auto-Curate system to reduce an estimated 400 million frames to 300,000 high-value annotated training assets that include action intent and context. In Phase 2, it converted this real-world material into digital assets representing spaces, actions, and objects, enabling synthetic-data generation for varied or hazardous scenarios that are difficult to capture directly. In Phase 3, Superb AI plans to build on these data and simulation capabilities to help K-EXAONE better understand and operate in physical settings.

Metafarmers, an AgTech company developing AI-powered crop-harvesting robots, worked with Superb AI to create scalable training datasets for perception models that recognize seedlings and mature harvest targets. As agricultural robotics demand grows amid rural labor shortages, the company needed to move annotation work beyond its engineering team while maintaining accuracy in complex crop imagery, where overlapping leaves and stems can make object boundaries ambiguous. Superb AI established two annotation workflows, using bounding boxes for seedlings and instance segmentation for mature crops, and built on Metafarmers’ existing labeled data and guidelines to define consistent standards before production. The project provided training data for the robots and a repeatable external labeling pipeline, allowing Metafarmers’ engineers to focus on robot and model development while retaining final review capabilities through Superb Platform.

A video-based behavior analysis AI company used its own system to extract useful training frames from observational footage but faced a growing bottleneck in converting those frames into accurately labeled data for pose and behavior recognition. Because the work required both bounding boxes and detailed multi-joint skeleton keypoints, including rules for occluded body parts, annotation was labor-intensive and frequently required rework as experimental results changed labeling standards. By adopting Superb Platform and Superb AI Data Labeling Services, the company centralized annotation, quality management, and dataset updates, defining consistent skeleton structures and visibility attributes within projects while outsourcing high-volume labeling and review tasks. This enabled researchers to focus more on model development, supported dataset expansion across multiple model iterations, and demonstrated the importance of scalable operational processes for maintaining precise pose-estimation training data.

As EV adoption expands in South Korea, where more than 750,000 EVs were registered by May 2025 and 238 EV fires had been reported through 2024, the National Fire Research Institute and Superb AI have collaborated to improve early fire detection in underground parking facilities. Their four-phase research program addresses the difficulty of identifying faint smoke from lithium-ion battery thermal runaway, especially in confined environments where delayed detection can worsen damage and firefighting access is limited. Using roughly 100,000 images from real incidents, controlled experiments, and generative-AI-created synthetic scenarios, the team trained and evaluated models specialized for smoke emerging beneath EVs rather than generic visible-fire detection. The selected model operates in real time on low-cost edge devices connected to existing CCTV systems, providing local web-based monitoring, visual and audible alerts, and event logging without requiring cloud servers. Field demonstrations and a full-scale underground smoke simulation validated the system under realistic conditions, and reported results indicate an 80% reduction in false alarms and a 65% shorter average detection time than an existing system. The final project phase focuses on deployment readiness and establishing quantitative standards such as detection distance and response time for future EV fire detection systems.

Superb AI has introduced Superb Academy, a multilingual learning hub featuring 26 short videos that cover the full data workflow in its platform, including project setup, collaborative labeling, Auto-Label, data curation, and data exploration. Organized into six chapters aligned with real workflows, the approximately four-minute videos can be viewed sequentially or accessed by topic, with transcripts, timelines, related user guides, and Korean-language Superb Recipe articles providing additional support. Available in Korean, English, and Japanese, the Academy is intended to help global teams onboard consistently and apply platform features to their own visual-data projects. Superb AI, founded in 2018, positions its enterprise training-data platform as a combination of automation, collaboration, and modular tools designed to accelerate the creation of high-quality machine learning datasets.