A personal goal – creating Intelligent AI that thinks and acts like humans

Junya FujimotoArtificial Intelligence Laboratory

Article｜2024-04-18

A vision of Intelligent AI that emulates humans

During university, I was involved in a laboratory project to develop a robot that assists with kitchen tasks (*1). In this project, we did manage to get the robot, via sensors, to recognize what were dishes and successfully transport them to the dishwasher. But there were also many things that we couldn't actually accomplish. For example, the robot struggled to handle dishes that were turned over or were a different shape. Additionally, to make it truly useful, a robot such as this needs to consider multiple factors such as whether the dishes are for a post-meal cleanup, where to store them after washing, and whether delicate items need careful handling. What may seem like simple instructions to humans can be challenging for robots to handle. Even just making the robot perform simple actions requires detailed programming inputs. During our project we realized that there were many experiences we could only gain through experimentation.

As a result of this work, I came to appreciate the remarkable adaptability of humans and developed a strong desire to develop an Intelligent AI capable of understanding the surrounding context, individual household rules, common sense and manners, allowing it to store dishes properly on shelves without needing explicit instructions. I am particularly grateful to my supervising professor who provided advice on experiments from various perspectives during my university years, which were full of learning experiences. The professor emphasized the importance of not only thinking but also using one's hands to create things and verify them, as well as to remember to enjoy it all!

A new approach to streamline the analysis of different working skill levels

Since joining the company, I have accumulated a lot of experience through many projects, working directly through trial and error to achieve fast results. One project that made the greatest impression on me was a joint development initiative for work segmentation AI technology used in the "OTRS+AI" service of Broadleaf Co., Ltd. (*2). In settings such as factories, variations in tasks due to the skill levels of workers pose a challenge. The existing technique of using cameras to record and analyze the tasks performed by beginners and skilled workers, identifying differences and iterating improvements, requires considerable time and effort. To shorten the training period for improving task proficiency, we took on the challenge of developing a new technology that could identify differences in tasks efficiently between beginners and skilled workers.

Typically, constructing an AI model that can accurately detect elemental tasks in a series of routine operations requires preparing a large amount of labeled training data, and associating video footage depicting the tasks with each elemental task manually. This helps prevent overfitting to the learned data. However, manually creating a large volume of training data is inefficient, even if one doesn't program the detection method. Our approach was to develop a model capable of recognizing tasks using only one set of training data for a single task. Although existing models did not support this, through trial and error and by incorporating the structures extracted from unsupervised analysis into probabilistic models, we were able to develop a proprietary model capable of recognizing tasks using only one set of training data for a single task (*3).

During the development process, I experienced the joy and excitement of seeing things I had conceived in my mind actually work well. And thanks to our success in enabling the differences in task times and movements for each elemental task between beginners and skilled workers to be confirmed easily, we received highly satisfactory feedback from our customers. Real-world scenarios present diversity and uncertainty, making productizing AI challenging. I believe it's essential to prioritize speed and practical usability over simply pursuing accuracy to an excessive degree. It's crucial to output quickly, under certain conditions, and trial the product in actual settings. I'm truly delighted that we were able to deliver the research results of this project swiftly to the world.

Enjoy each fulfilling day

My day starts with breakfast. Since I mostly work remotely, I begin my workday after dropping off my child at the nursery. I believe it's important to allocate enough time for both hands-on work and thinking, so I strive to maintain a balance between meetings and actual tasks. Lately, during my free time, I've been reading specialized books on machine learning. I'm not only studying the theory but also consciously thinking about how to apply new AI technologies. Of course, work doesn't always go smoothly. There have been times when I faced difficulties or periods when things didn't go well. During those times, I turn to music! The lively and energetic performances of Momoiro Clover Z's songs and dances cheer me up and give me the strength to overcome difficulties.

I enjoy sports, and I've been playing badminton as a hobby for over 20 years. Every year, I practice once or twice a week for badminton tournaments. Some of my practice partners are members of Fujitsu Research, and I value our connections outside of research. Apart from badminton, watching TV with my wife or going for a bike ride together is also great for refreshing.

Memories of my father, sparking a love for R&D

When I was young, I used to enjoy playing games on the Family Computer (Famicom). I remember a time when the game screen went black due to a fault in the Famicom's cable connector. My father, who worked in R&D, dismantled the Famicom, soldered the cable to the circuit board, and repaired it for me. Seeing the green circuit board inside the Famicom for the first time was impressive, and I admired how things could be repaired so quickly. I think my father's influence played a big part in my decision to pursue a career in R&D. Although our research fields are different, the desire to create technology that benefits society is something we share as parent and child. While AI applications in business, like ChatGPT, have seen remarkable advancements, there are still many challenges to overcome for Intelligent AI and robots to support people's daily lives. Improving AI's ability to understand the real world is one of these challenges. To contribute to solving these challenges and advancing technology, I want to continue R&D with my own unique approach in the future.

Editor's note

Junya told me that he has a deeper interest in the intelligence that controls robots rather than its hardware. Much of his inspiration comes while taking walks or bathing. It seems that he often writes down research ideas, sleeps on them overnight, and then reconsiders them. We are looking forward to the day when the artificial general intelligence that he wants to achieve will arrive. (Xiang Yi Peck, Technology Strategy Unit)