[Editor's Note] MDS Tech is a technology company that supports embedded system development in various industrial sectors, including automotive, defense, aerospace, and semiconductors, and is enhancing the developer environment through development and verification solutions, including TRACE32. Based on a partnership with Arm, the company provides solutions covering the entire V-model lifecycle, such as compilers, debuggers, and dynamic testing tools, going beyond simple supply to support education, technical support, and verification. In particular, for physical AI development, it supports the entire process of platform development and verification using Jetson, Keil MDK, TRACE32, and DT+Trace. Against this backdrop, MDS Tech is scheduled to give a presentation on the topic of "Development Strategies and Practical Tips for Physical AI Development" at the "e4ds Physical AI Frontier 2026," to be held on March 6. Accordingly, this publication met with Manager Kim Min-ki of MDS Tech, who was in charge of the presentation, to hear about the key points for designing physical AI.


■ Please introduce MDS Tech.

MDS Tech is continuously striving to lead technological advancement and innovation in development fields across various industries, including automotive, defense, aerospace, and semiconductors.

TRACE32 is a hardware debugger that has been the standard for embedded system development for 30 years. It is supported in all environments and possesses unique technology applicable to every project.

MDS Tech provides optimal solutions to customers through a partnership with Arm, a global semiconductor IP technology company.

MDS Tech is constantly taking on challenges to lead the growth of domestic industries and contribute to Korea's smart future.

■ I would like to hear an introduction to MDS Tech's key products and how MDS Tech contributes to developers.

MDS Tech plays a key role in the V model, a software development cycle.

We sell development kits, compilers necessary for software development, debuggers for code debugging, and dynamic testing tools for integration testing.

Notably, we serve as the domestic distributor for Nvidia platforms essential for artificial intelligence computing, such as Jetson Nano, Orin, and Thor, as well as the hardware debugger TRACE32, Arm’s Keil MDK and Arm DS required for code compilation, and probe code-based dynamic testing. There is a tool called DT+Trace.

MDS Tech is actively helping to solve problems, rather than simply supplying products.

It provides optimized tools and technical support so that developers do not get bogged down in complex hardware control or compatibility issues.

We provide a variety of services, including training, technical support, and verification for newly introduced solutions.

Through this, MDS Tech's technology philosophy is to create an environment where developers can focus solely on the essence of the service and the implementation of innovative algorithms.

■ The key elements that must be considered for physical AI to operate stably in real-world environments are

Unlike general AI, physical AI has a 'physical entity'.

The physical entity can be a robot or a drone, as shown at the CES 2026 exhibition.

This means that the hardware entity interacts with the user every moment, and the important thing is the real-time nature of the electronic system hardware.

When dealing with real-time capabilities, the two most important things are time analysis and safety regarding functionality.

Time analysis is essential for rapid and accurate analysis in environments where sensors or devices are integrated, and it is a verification method necessary to prevent unintended accidents.

This also requires an analysis of how efficiently limited power and memory resources are being used.

The second is functional safety.

There are too many variables in the user environment where software is operated.

Therefore, standard specifications necessary for software development, such as ISO, must be defined, and development must be carried out in compliance with those specifications.
Safety certification, exception handling, and risk analysis must be incorporated from the design stage to ensure the system maintains a 'safe state'.

■ What are the most frequent problems that occur during the initial design phase of physical AI system development, and what are the key points developers should pay attention to?

The most common problem is the failure of performance prediction.

This is a case where an AI model that ran perfectly on a high-spec workstation fails to perform well or freezes due to insufficient memory when deployed on an actual target board.

The key point developers need to pay attention to is target-centric design.

Instead of optimizing after the design is complete, you must consider the specifications of the MCU or GPU to be used from the beginning to adjust the number of model parameters or establish a plan for utilizing the accelerator (NPU).

Furthermore, the key to reducing trial and error is to introduce a process that actively utilizes virtual environments to verify software logic without hardware during the early stages when it is difficult to secure sufficient data from the actual physical environment.

I am curious about the typical types of risks involved in physical AI development and the approaches taken during the development phase to mitigate these risks.

The biggest risk is 'abnormal behavior regarding edge cases'.

The biggest risk for developers and users is that it is impossible to know what unexpected behavior the sensor or controller will exhibit when it encounters a situation not found during the data learning phase.

To manage these risks, various tests are conducted during the development and verification phases.

Among them, a commonly used method is to identify risks during the code development phase.all.

Analyze trace information based on the source code running in the controller module, or check memory usage and CPU utilization.

With the controller integrated, real-time performance can be verified and potential edge cases predicted by analyzing the execution time per task.

In cases where there is no development platform or target, dynamic verification can be performed based on a virtual environment.

This method involves conducting tests in a virtualized environment before the actual hardware is released, testing the responsiveness of the AI and controller by injecting user-generated scenarios into a digital twin environment without sending the car to the road.

■ MDS Tech's solution for physical AI development is

MDS Tech supports the entire process of 'platform development and verification'.

First, the Nvidia Jetson Series is designed for image and video processing and the execution of artificial intelligence applications based on high-performance GPUs; it provides high performance even in a compact size and weight, and enables the development of edge-based applications.

Arm Keil MDK v6 provides a development environment optimized for device development.

There are compilers and IDEs such as Keil MDK and Arm DS, and using a virtual hardware environment called FVP helps you start advanced development even without a real physical board.

Next, when debugging source code or development code is required, the developer can use the TRACE32 hardware debugger to check the current operating status of the device.

You can monitor variable values or conduct additional tests according to pre-written test cases when a function call occurs.

In the dynamic testing phaseIt is possible to analyze code coverage and execution time using DT+Trace.

Is there a message or insight you hope participants will take away from the upcoming Physical AI Frontier 2026?

I think 2026 is a time to see how existing AI can be applied to our lives.

I hope the developers participating in this event will have the opportunity to experience the various physical AI platforms available to them.

I hope you gain practical engineering methodologies that go beyond simply knowing "which AI model is good," covering how to apply the model to actual hardware equipment, how to perform testing and verification, and how to maintain it.

We hope you will explore the various case studies and solutions prepared by MDS Tech to see how you can realize your ideas in a realistic and safer way.

■ Finally, is there anything you would like to say to the readers?

The development and verification of physical AI is by no means an easy journey.

There are edge cases that cannot be reproduced, or cases where data for training and testing does not exist.

Does the absence of input mean the output cannot be verified? Not necessarily.

There is also a method to simultaneously verify hardware and software by virtualizing the target environment or by having the user directly input input values via a hardware cable.

In some cases, all information generated by the target can be logged or stored for analysis.

MDS Tech possesses years of accumulated industry-specific know-how, solutions, and trust.

femaleMDS Tech will be with you in the physical domain to help your innovative ideas create value beyond AI models.