Config sells robotics data infrastructure and model training services to robotics labs, manufacturers, and enterprises that need to teach robots to manipulate objects with two arms. The buyer's problem is that unlike language models, which train on freely available internet text, useful manipulation data often requires someone to perform a task, record it, and convert it into a format a specific robot can learn from. Config handles that entire chain.

Catalog is Config's outright-purchase data product: pre-built datasets of human and robot manipulation recordings, browsable by task, gripper or hand type, and source. The current library lists more than 200,000 total hours of human and robot data across more than 10,000 tasks, collected primarily at Config's Hanoi studio and Seoul lab. Teams training their own models can buy recordings in LeRobot format, a standard used by robot learning tools, without operating a collection pipeline. Separately, Config released HABIT on August 7, 2026, a 164-hour, 60-task human-robot interaction dataset under CC BY 4.0, publicly available on Hugging Face. The public research release provides examples of collaboration, side-by-side work, and following human gestures.

Loop is the platform a customer uses when they need Config to collect, convert, train, and deploy a policy, the model that chooses a robot’s movements, for a specific task. A human operator demonstrates the task—folding cloth, sorting recyclables, serving food—while multi-camera rigs record hand motion and workspace context. Config's curation team reviews each episode for quality. The Conversion Engine then translates the human demonstration into robot-ready trajectories matched to the customer's specific gripper and arm, so a human’s recorded movements can supervise the robot’s different body. This converted data, or direct teleop data from Config's Seoul lab, feeds into the Policy Engine, which fine-tunes CFG-1 or an open-source model such as Physical Intelligence's Pi 0.5 or Nvidia's GR00T N1.6 to the target task. The trained policy runs on the robot under live inference, and every rollout is scored and fed back as new training data for the next cycle.

CFG-1 is Config's own 2B-parameter foundation model for bimanual manipulation, trained on 10K+ hours of in-house data. The company claims sub-50ms inference on a single RTX 5090 and 80%+ task success after one hour of task-specific teleop fine-tuning. These are Config's reported benchmark results, not industrial reliability guarantees. CFG-1's three-minute native memory window lets it handle multi-step tasks—pick, place, recover—without losing context, while a sliding window of recent frames drives fine-grained action prediction.

Config earns revenue through two channels. Catalog sales are outright data purchases: a robotics lab buys a pre-converted dataset by task and embodiment. Loop engagements are negotiated contracts in which Config collects custom data, converts it, trains a model, and deploys it on the customer's hardware—scoped to whatever subset of the six modules the customer needs. Current customers include large manufacturers, system integrators, and companies in agriculture and defense.

Human labor is a substantial delivery cost. Config’s current site describes more than 100 operators in Hanoi producing approximately 20,000 hours of human demonstration data per month; its earlier May profile described a roughly 300-person data workforce across its operations. QA, conversion compute, GPU training, and on-site integration add further costs. The economic advantage Config is building is reusability: once a human demonstration is converted for one robot embodiment, the same episode can be re-converted for additional embodiments, spreading collection and curation costs across multiple customers. That advantage is conditional—it holds only if enough customers share overlapping task types and if converted data transfers well across hardware platforms. Bespoke tasks with unique fixtures or objects offer less opportunity to reuse the original collection work.

Config competes with Lightwheel and Maniformer for robot training data budgets, while NVIDIA and Physical Intelligence supply models that customers can build around themselves. Buyers must decide whether to operate their own collection and training systems or pay a supplier to deliver data and working task models.

Lightwheel is a Physical AI infrastructure company offering simulation-ready 3D assets, egocentric human data collection via its EgoSuite devices, and the RoboFinals evaluation platform. Lightwheel combines human demonstration collection across several countries with physically calibrated 3D scenes and simulation testing. Where Config converts human demonstrations into robot-native trajectories and trains policies end-to-end, Lightwheel emphasizes the simulation layer: reconstructing physical environments digitally, generating synthetic training data, and evaluating policies in simulation before real-world deployment. Config's differentiation is the closed loop from real-world human data through conversion to deployed policy; Lightwheel's is the sim-to-real pipeline and its NVIDIA ecosystem integration.

Maniformer, based in Shanghai, operates a catalog of task datasets, custom collection services, and capture hardware. Its MEgo hardware—a precision gripper and head-mounted camera system—competes directly with Config's collection rigs. Maniformer's Hive Data Co-creation Initiative positions it as a marketplace connecting data providers and consumers globally, whereas Config controls its own collection workforce and conversion pipeline. Config’s own collection workforce lets it manage quality directly, while Maniformer is trying to expand supply through a wider network of partners. Either approach must prove that the data improves a buyer’s robot on its actual tasks.

Nvidia's GR00T and Physical Intelligence's Pi 0.5 are the open foundation models Config explicitly supports inside Loop's Policy Engine. A buyer can use these models with its own demonstrations instead of commissioning Config’s full service. Supporting several models helps Config sell collection, conversion, and deployment even when a customer prefers another model; open tools also reduce the amount of proprietary software buyers need to purchase.

Config’s Catalog addresses robotics teams buying training data, while Loop provides a route into the larger budgets used to automate physical work. Expansion depends on selling reliable task execution to operating businesses, rather than only supplying recordings to research teams.

Config's strategic investors—Samsung, Hyundai, LG—are themselves large-scale manufacturers with active robotics programs. The Loop platform's full pipeline, from task demonstration through deployed policy, is designed for exactly this buyer: a factory operator who wants a bimanual robot to perform a specific assembly, sorting, or packaging task without building an in-house data team. Config's Samsung C Lab Outside selection gave it hands-on experience producing data for real industrial sites. The new budget here is not a dataset purchase but a deployment contract: Config could sell site-specific data collection, training, and deployment services, with follow-on work as production requirements change. The dependency is proving that 80%+ benchmark success translates to industrial-grade reliability on factory floors with variable lighting, part tolerances, and throughput requirements.

HABIT's 60 tasks span table service, food storage, shelf cleaning, donut serving, and waste sorting—all performed with a human sharing the workspace. These controlled research tasks resemble work performed in hospitality and other service settings. A hotel, hospital, or food-service operator that wants a robot to clear tables or sort waste alongside staff represents a buyer whose budget currently goes to manual labor, not to robotics data vendors. Config would earn revenue by licensing task-specific policies trained on HABIT-style data and providing ongoing Loop deployments tuned to each site's layout and objects. The dependency is commercial: HABIT demonstrates the capability in controlled settings, but no commercial service deployment has been announced, and the gap between research demonstration and a paying hospitality customer is substantial.

Transfer reliability: Config’s reported task success rates do not establish reliable performance across different factories, objects, and robot hardware. If converted demonstrations transfer poorly, customers need more site-specific collection and Config loses part of its reuse advantage.

Collection costs: Config operates human data studios and plans to expand the amount of data it supplies. If new tasks require substantially more operators and quality checks, higher sales may not produce the cost improvements expected from reusable software and datasets.

Buyer insourcing: Loop supports open models from NVIDIA and Physical Intelligence, giving customers alternatives to CFG-1. Buyers that combine those models with internal data teams could reduce their use of Config to a smaller collection or conversion contract.