🎙️ If you're considering your career path after completing your undergraduate degree or graduate school, be sure to check out this interview!

Today's team interview features the "Perception Cell," which handles the video data processing from the cameras, essentially the "eyes" of the delivery robot, Neubie.

The members of the Perception Cell, Jeongbae and Maro, share the common experience of pursuing graduate school after completing their undergraduate studies.

We asked them a few questions about their work and their decision to attend graduate school.

✏️ Here’s what you can discover in Perception team's interview:

Q. We have two team members with us for this interview.

     Could you each introduce yourselves briefly?

Jeongbae : Hello, I'm Jeongbae Son, and I’m responsible for deep learning model research and development in the Perception Cell of the Neubility Autonomy Team.

Maro : I'm Maro Jeon, and along with Jeongbae, I work on the task of lightweighting deep learning models in the Perception Cell.

Chapter 1. Perception Cell?

Q. What does the Perception Cell do?

Jeongbae : The Perception Cell is the team responsible for the very first stage of autonomous driving. Simply put, it plays the role of the "eyes" of Neubie. Just like how we need to see with our eyes to decide our next actions, the AI in the Perception team processes the information coming from the camera. Once this information is processed, the next team can use it to make the vehicle move. As the name "Perception" suggests, the core role of this team is to recognize the surroundings, identifying paths and obstacles to navigate through.

The autonomous driving system is broadly divided into three main parts: Perception, Localization, and Planning & Control. Perception focuses on detecting the surroundings, identifying accessible areas and obstacles. Localization determines the vehicle's current position and the destination. Planning & Control then uses this information to develop the optimal driving strategy.

Q. What technologies are used?

Jeongbae : At Neubility, the perception system primarily uses cameras and deep learning models to recognize the surroundings. Specifically, the deep learning models include Semantic Segmentation (to classify the terrain the robot will pass through), Object Detection (to identify obstacles), and Depth Estimation (to reconstruct 2D images into 3D space).

Additionally, in order to advance the technology, it's essential to quickly accumulate data. To accelerate this process, we use techniques such as Domain Adaptation and Auto-labeling. These are just a few of the various technologies we use.

Jeongbae Son (Autonomy Perception Cell) 

Q. It’s easy to understand that you play the role of our eyes!

     I’d love to hear more about what the Perception team has been focusing on lately.

Jeongbae : We've recently been working on ensuring that Neubie can quickly adapt to new environments. Since we’re a startup, we often need to actively take on new business opportunities, no matter where they arise. As a result, the robots we develop often face unfamiliar and new environments. However, typical deep learning models don’t perform well in new environments, as they differ from humans. This sometimes causes failures or issues with performance.

However, for Neubility to expand its business areas, it's crucial to quickly respond to new environments and the issues that arise. That's why we follow the latest research trends in areas like domain adaptation and internalize external knowledge into Neubility's proprietary technology, enabling Neubie to handle problems when it encounters new environments.

In the past, it used to take more than three months to develop a deep learning model suitable for a new environment in order for Neubie to operate there. But recently, we've made such progress that we can complete the optimization for operating in a new location within a week.

Q. What are some examples of new elements Neubie needs to adapt to?

Jeongbae : For example, a deep learning model trained to operate on asphalt or pavement might not work well when it’s in a golf course environment. A golf course has grass on either side of the paths, and the types of obstacles are different from those in the city. Similarly, when Neubie, which operates outdoors, moves indoors, new issues arise. The floor may be shiny, causing reflections of other objects, or there could be a higher concentration of people. We focus on helping Neubie quickly adapt to such changes.

Q. I heard that you recently joined the Perception team, Maro.

     I'm curious about the role you play.

Maro: When Jeongbae and the other team members develop deep learning models, my role is to make those models run faster and lighter. For a deep learning model to work smoothly on a robot, it needs to be lightweight.

For example, Neubie has five camera eyes, and processing the information from all five can be quite demanding. So, my job is to make the deep learning model lightweight enough to handle the processing of all five pieces of information.

Maro Jeon (Autonomy Perception Cell) 

Q. Let me start with this question: If you could go back in time, would you choose

     to attend graduate school again or not?

Jeongbae: Absolutely, I would.

Maro: Without a doubt, I’d go again.

Q. What made you decide to pursue graduate school before joining the company?

     Could you share any concerns you had and how you resolved them?

Jeongbae : Actually, my undergraduate major wasn’t related to artificial intelligence at all—it was in materials engineering, specifically focused on semiconductors. During my senior year, as I was preparing for job applications, I came across the concepts of artificial intelligence and deep learning by chance. Around that time, I watched The Imitation Game, which introduced me to the fascinating world of computer vision. That experience left a strong impression on me and ultimately led me to pursue a master’s degree in a field completely different from my undergraduate studies.

At the time, many of my peers who studied materials engineering had already secured jobs, so I had my share of doubts. I often wondered whether it was the right decision to give up on the field I had dedicated four years of undergraduate study to. Even after enrolling in graduate school, competing in a completely unfamiliar field was incredibly challenging. However, I have no regrets about that choice.

What makes the autonomous driving field so rewarding is the tangible results of my efforts. Whenever I solve a problem or develop a new feature that wasn’t functional before, and then see it working in real life, it brings me immense satisfaction. That immediate sense of achievement is one of the greatest joys of being an engineer.

Maro : I majored in naval architecture in my undergraduate studies, but at the time, I felt it was a somewhat traditional and outdated field. With self-driving cars becoming a reality, I thought creating autonomous ships would be an exciting challenge, so I started looking for a research lab in that area. Back then, I had only about three days to decide whether to pursue graduate school. Without overthinking it, I decided to join the lab—and I’ve never regretted that choice.

Graduate school was where I began researching AI applications in ships, and it laid the foundation for my career path in artificial intelligence.

Q. Both of you seem to have clarified your career paths through graduate school.

     You mentioned you don’t regret the decision—what specifically about the master’s

     program were you satisfied with?

Jeongbae: During graduate school, I delved deeply into the mathematical concepts behind how AI models operate. While not everything I researched directly connects to my current work, the strong foundation I built back then has been incredibly beneficial. For example, my ability to review academic papers and adapt those technologies to our needs is a skill I developed during my master’s studies.

More importantly, since my undergraduate degree wasn’t in AI, pursuing a master’s degree was crucial for breaking into the field I was passionate about. At the time, it might have felt like taking a longer route, but being able to study what I wanted in graduate school and now working in that field is something I’m deeply satisfied with.

Maro : The AI I studied during my master’s program differs significantly from the AI I work with now. So, while the specific skills and knowledge I gained back then may not directly apply, the process of reading research papers and solving various problems has been incredibly helpful. The logical approach I developed while discussing ideas with my professor and determining clear directions is also essential for my current work. At Neubility and other startups, much of the work revolves around solving previously unsolved problems, and the logical thinking I honed during graduate school has been invaluable.

Most importantly, pursuing a master’s degree in AI can elevate the level of companies you can join. Simply put, it gives you a competitive edge, making it easier to secure positions at more prestigious companies compared to someone with just an undergraduate degree. That’s another reason I highly recommend pursuing a master’s program.

Q. This interview will likely be read by those who are considering pursuing graduate

     school or thinking about joining Neubility after completing their master's degree.

     Do you have any advice for them?

Jeongbae : The biggest difference I felt between academia and industry is "application." In research labs, the goal is often to improve performance metrics and write impressive papers. However, in the real world, where we need to deliver commercialized services, the ability to quickly adopt and apply technology is far more important.

While writing great papers is important, in practice, it’s much more crucial to be able to quickly implement technology and collaborate with others to solve problems. So, my advice would be to focus on building a strong foundation in the fundamentals and programming skills, rather than becoming too fixated on publishing papers. Concentrating on the process of solving problems together with a team is key.

Maro : I feel the same way. If you're pursuing a master's degree, I would suggest focusing on writing your thesis according to the requirements and giving it your best effort. For those about to start their professional careers, I recommend taking the time to think carefully about how you want to shape your career path. Also, it’s crucial to clearly know what you want to do.

Chapter 3. Tips for Success in Applying to Neubility

Q. What kind of technical experience and background does Neubility value?

     Can you share any tips for success in the application process?

Jeongbae : I believe that having a mindset to persistently solve problems without giving up when a challenge arises is crucial. Unlike autonomous driving in vehicles, our robot needs to navigate crowded sidewalks, which presents many issues to address. There are no clear answers, and unpredictable problems arise every week and every day.

For example, while it's easy for a person to cross a crosswalk by following traffic lights, our robot needs to consider things like "which traffic light to follow" and "how to navigate safely among many people." There are many factors to pay attention to. As a result, the destination for solving these problems may not always be clear. Nevertheless, if you have a mindset that doesn't give up, you can say that you're halfway prepared.

Maro : Above all, I believe it's important to understand what you're working on in greater depth. It's not enough to just have tried what others are doing. Instead, you need to really grasp the difficulties you've faced, the efforts you've made to overcome them, and why you chose a particular algorithm. During interviews, it’s helpful to show what you know. The interviewer will likely ask more in-depth questions, and to answer them well, you’ll need to have a thorough understanding. (Laughs)

The difference between knowing something on the surface and deeply understanding it comes down to whether you've just read about it and followed along, or if you've applied that technology in other contexts. When you try it in a different environment, new challenges inevitably arise. Figuring out how to solve those problems will likely help you improve significantly.

Q. What goals does the Perception team have for the future?

Jeongbae : To start with the goal of Neubility as a company is to make short-distance deliveries easily accessible to everyone in South Korea through autonomous delivery robots. In order for many people to use this service, we need to reduce costs, which means finding alternatives to expensive sensors. The Perception team is working on advancing technology based on cameras to achieve this.

The goal of the Perception team is to enable the robot to perceive the world solely through a camera, just as humans perceive the world with their eyes. Specifically, we are developing technology that converts 2D images captured by the camera into 3D representations. By using camera-based technology, we aim to lower the cost of the robots, making them accessible to more people while ensuring they function effectively.

Q. Lastly, what kind of person would you like to work with on the Perception team?

Jeongbae : The Perception team develops functions that allow Neubie to perceive the information necessary for autonomous driving, such as "road surface recognition" and object avoidance. Since we use cameras, it would be great if the person has a solid understanding of cameras, including concepts like intrinsic and extrinsic parameters and the coordinate systems that define the camera. Our technology is based on AI deep learning, so familiarity with model structures related to computer vision and the ability to customize them in code would be ideal.

Also, as mentioned earlier, we are looking for someone who can independently analyze and solve problems and has an active mindset toward problem-solving.