Physical AI is moving from research labs into humanoid robots, autonomous machines and industrial systems, but the technology faces a fundamental bottleneck: training data that accurately represents the physical world. NAVER D2SF is making its third investment in South Korean startup NdotLight, betting that the company’s simulation-ready 3D data infrastructure can help address that challenge.
NAVER D2SF Makes Third Bet on Physical AI Data Startup NdotLight
NAVER’s corporate venture capital arm, NAVER D2SF, has made a follow-on investment in NdotLight, a company developing data infrastructure for physical AI.
The KRW 15 billion funding round was led by Korea Development Bank, with NAVER D2SF participating for the third time. Its previous investments came during NdotLight’s Pre-Series A round in 2021 and Series A financing in 2022.
The latest investment reflects a growing realization across the robotics industry that AI models alone are not enough to produce capable physical machines.
Robots need to understand environments, objects, motion, friction, mass and physical constraints. Generating enough high-quality data to teach those concepts is expensive and time-consuming.
NdotLight is positioning its TRINIX platform as part of the infrastructure layer intended to solve that problem.
The Data Problem Behind Physical AI
The rise of physical AI has created a new version of the data challenge that has already reshaped generative AI.
Large language models require enormous quantities of text and multimodal data. Physical AI systems need data that captures how objects and environments behave in the real world.
One approach has been teleoperation, in which humans control robots while their actions are recorded as training data. Another is collecting data directly from robots operating in physical environments.
Both approaches have limitations.
Real-world collection can be slow and expensive, while datasets may lack the detailed physical information required by simulation environments.
That creates an important gap between data that shows what happened and data that enables a simulation to reproduce why it happened.
NdotLight’s approach is aimed at the latter.
TRINIX Turns 3D Assets Into Simulation Data
The company’s TRINIX platform generates what NdotLight describes as “simulation-ready” 3D data.
Instead of creating only visual representations of objects, its automated pipeline generates 3D assets containing physical properties such as mass and friction, along with joint structures and collision boundaries.
Those details are critical for robot simulation.
A robot learning to grasp an object needs more than an image of the object. The simulation needs to understand where its surfaces are, how much the object weighs, how it interacts with other objects and how forces affect movement.
That makes NdotLight’s technology less comparable to conventional 3D content creation software and more relevant to the emerging synthetic data and simulation infrastructure market.
TRINIX also integrates with NVIDIA Omniverse, NVIDIA’s platform for industrial simulation and 3D collaboration. That integration allows NdotLight to target large-scale datasets designed for simulation-based physical AI training.
Simulation Is Becoming a Core AI Training Layer
Simulation is increasingly important as robotics companies attempt to scale training without relying entirely on physical machines.
A simulated environment can allow robots to repeat tasks thousands or millions of times without the wear, safety risks and operating costs associated with physical testing.
The challenge is simulation quality.
If virtual environments fail to represent real-world physics accurately, models trained inside them may not transfer effectively to physical robots. This is commonly described as the sim-to-real gap.
Better simulation-ready data can help narrow that gap by giving models more realistic representations of objects, environments and physical interactions.
NdotLight is targeting this infrastructure layer rather than building a robot or a foundation model itself.
That puts the company in an increasingly important part of the physical AI stack: the data between the physical world and the AI model.
Early Customers Span Humanoid Robotics and Industrial AI
NdotLight says it currently supplies training data to humanoid robotics and robotics foundation-model companies including Holiday Robotics, AeiRobot, ROBROS and RLWRLD.
The company is also participating as a data supplier in physical AI initiatives involving major industrial companies including Hyundai Motor Company and LG Electronics.
Those relationships are significant because they suggest demand for physical AI data is extending beyond startups and research projects into industrial applications.
Automotive manufacturers, electronics companies and robotics developers are all exploring ways to use AI for physical tasks. That creates potential demand for standardized, scalable data infrastructure that can support different simulation environments and robotic platforms.
NAVER’s Continued Support Signals Longer-Term Strategy
The investment also illustrates NAVER D2SF’s approach to venture investing.
Rather than exiting after an early financing round, the corporate venture arm has continued to back NdotLight as the company’s market focus evolved.
According to NAVER, its relationship with NdotLight began when the startup was a resident company at NAVER 1784, NAVER’s technology-focused headquarters, where the organizations collaborated around 3D content creation.
The latest investment is intended to open additional opportunities for collaboration around physical AI.
That strategic relationship could become more important as NAVER explores AI beyond conventional digital services and into embodied systems and physical-world applications.
A Different Kind of AI Infrastructure Race
The physical AI market is attracting investment across several layers.
Companies such as NVIDIA are supplying the compute and simulation infrastructure. Robotics companies are developing foundation models and hardware. Industrial companies are deploying robots into manufacturing and logistics environments.
The missing layer is increasingly high-quality physical-world data.
NdotLight’s bet is that generating this data programmatically can scale more efficiently than collecting everything from physical robots.
The opportunity is substantial, but so are the technical challenges.
Synthetic data must accurately represent the physical world. Simulation environments must translate into real-world performance. And enterprises need data pipelines capable of supporting different robotics platforms and training frameworks.
NdotLight’s integration with NVIDIA Omniverse and its work with robotics and industrial customers give the company an entry point into that emerging ecosystem.
Physical AI May Need a Data Infrastructure Industry of Its Own
The generative AI boom created an enormous infrastructure market around GPUs, cloud computing, model training and data pipelines.
Physical AI could develop a comparable stack, although its requirements are different.
Robotics needs 3D environments, physics-aware assets, sensor data, simulation, teleoperation data and real-world feedback loops.
That means companies building the data infrastructure underneath physical AI could become strategically important even if their products are invisible to the end user.
NAVER D2SF’s third investment in NdotLight is a signal that investors are beginning to recognize this layer as a potential bottleneck—and an opportunity.
If physical AI is to move from controlled demonstrations to large-scale deployment, robots will need far more than better models. They will need better representations of the world they are expected to operate in.
That is the market NdotLight is trying to build around.
Market Landscape
The physical AI ecosystem is developing around several interconnected layers:
| Layer | Role |
|---|---|
| Compute | GPUs and AI accelerators for model training and simulation |
| Foundation models | General-purpose models for robotics and embodied intelligence |
| Simulation | Virtual environments for scalable robot training |
| 3D data | Objects, environments and physics-aware assets |
| Robotics hardware | Humanoid, industrial, mobile and collaborative robots |
| Real-world data | Teleoperation, sensor and deployment feedback |
| Industrial applications | Manufacturing, logistics, automotive and electronics |
NVIDIA’s Omniverse is becoming an important part of the simulation ecosystem, while robotics companies are competing to develop foundation models that can generalize across tasks and environments.
NdotLight is targeting the 3D data and simulation layer, where accurate physical properties can determine how useful synthetic training environments ultimately become.
For enterprise adopters, the strategic question is whether simulation can reduce the cost and time required to train and validate physical AI systems while maintaining sufficient real-world transfer.
Top Insights
- NAVER D2SF’s third investment in NdotLight signals continued confidence in simulation-ready 3D data as a critical physical AI infrastructure layer.
- NdotLight’s TRINIX platform generates 3D assets with physical properties, joint structures and collision information designed for robotics simulation and training.
- Integration with NVIDIA Omniverse positions NdotLight within the growing ecosystem connecting synthetic data, simulation and robotics foundation models.
- Customers and projects involving humanoid robotics, Hyundai Motor Company and LG Electronics indicate demand extending from startups into industrial physical AI applications.
- Physical AI may require its own infrastructure stack spanning simulation, 3D data, robotics models, compute and real-world feedback as deployment scales.
Power Tomorrow’s Intelligence — Build It with TechEdgeAI
