Foxconn, renowned for its smart manufacturing expertise, is now significantly expanding its focus into the realm of smart healthcare. The company is demonstrating its integrated ecosystem strategy across four key medical implementation scenarios: clean rooms, operating rooms, consultation rooms, and nursing stations. We take an exclusive, firsthand look inside the colonoscopy AI Agent laboratory.

A pharmacist is preparing chemotherapy drugs for a cancer patient within a clean room. This seemingly routine scene represents one of the most hazardous tasks in a hospital, as the volatile cytotoxic drug molecules can pose a direct health risk to the pharmacist if mishandled. At Taipei Veterans General Hospital, where one-third of the patients are cancer cases, this process is performed hundreds of times daily.

At the COMPUTEX exhibition earlier this month, an AI solution for executing such dangerous tasks was showcased for the first time. This year's event expanded to include the Taipei International Convention Center, featuring a new AI robotics application zone. With the theme "AI Together," it signaled a shift from last year's individual displays to a collaborative effort in building an ecosystem.

In partnership with Taipei Veterans General Hospital, Foxconn is promoting the automation of chemotherapy drug preparation. As a leader in the electronics industry, the company is not only exhibiting smart manufacturing and smart city solutions but is also concentrating on smart healthcare, positioning itself as a key integrator within the ecosystem.

Foxconn's Sub-Group B Business Unit General Manager, Jiang Zhixiong, highlighted the four "D"s present in hospitals: Dirty, Dangerous, Dull, and Difficult. The preparation of chemotherapy agents is a quintessential Dangerous scenario.

To address this, Foxconn has designed an end-to-end automated dispensing workflow. It begins with the introduction of automated dispensing robots for chemotherapy and radioactive drugs, connects with AMRs for delivery, and utilizes the Nurabot equipped with biometric authentication to transport the medications to 12 outpatient satellite injection stations at Taipei Veterans General Hospital. This system is slated to commence clinical trials as early as the fourth quarter of 2026.

Unlike the orderly environment of a factory, hospitals are dynamic, facing challenges such as emergency situations, chaotic patient flow, and yet requiring uncompromising safety standards. This is precisely the motivation behind Foxconn's increased commitment to smart hospitals.

"If it can be implemented in a hospital, it can be implemented anywhere," stated Jiang Zhixiong. Moving beyond previous exhibitions of single robotic product breakthroughs, the current focus is on connecting points into lines and lines into planes. Beyond chemotherapy drug preparation, the exhibition also featured AMRs for transporting waste and kitchen scraps, as well as AI nursing robots for delivering specimens, medications, and patient education—tasks covering the 3D aspects, all deeply integrated with Taipei Veterans General Hospital. "We are integrating fragmented products into a cohesive workflow, emphasizing proximity to the field and practical application."

Shifting the scene to the operating room, Foxconn's scrub nurse robot provides a solution for the Difficult aspect.

The scrub nurse robot is a robotic arm designed to accurately pass surgical instruments to doctors in the operating theater. When it debuted at GTC in March, it could only recognize five types of instruments. After three months of training, its recognition capability has expanded to 18 instruments.

The initial five instruments learned in the first stage represent the five major categories of surgical tools. As the number increases, the differences between some instruments become subtle—variations in tip angle, thickness, or minor structural details. Combined with environmental factors like metal glare, blood contamination, and occlusion, the recognition difficulty multiplies.

By incorporating a third-person perspective camera, the AI has evolved from interpreting flat images to understanding three-dimensional space, comprehending the instrument's position, angle, and spatial relationships. The next goal is to complete learning the 36 instruments required for a relatively straightforward thyroid surgery, increase speed, and ultimately aim to expand recognition to 100 or 200 instruments.

To ensure the scrub nurse robot meets the time-critical demands of surgery, the Foxconn team personally observed doctors' every move during procedures. "Cutting the skin, drilling bone, I even saw bone marrow spray out," recalled Lin Qifan, User Experience Design Director for Digital Health at Foxconn's Sub-Group B Business Unit. She believes this immersion is essential to understand the subtle nuances of a doctor's usage.

Translating clinical observation into capability involves implementing a "Vision-Language-Action" end-to-end architecture. It uses vision to interpret space and objects, receives natural language commands from the doctor via bone conduction headphones, and finally converts these into actions. After model inference, it controls the robotic arm to rotate precisely, safely, and accurately handing the scalpel handle to the surgeon.

In collaboration with Kaohsiung Medical University Chung-Ho Memorial Hospital, Congtai Technology, Olympus distributor Yuanyou Industrial, and other ecosystem partners, Foxconn has co-developed the colonoscopy AI Agent named "CoDoctor Endovia."

"Even an experienced endoscopist performing several colonoscopies a day might miss something due to fatigue, decreased attention, or variations in lesions," pointed out Dr. Cai Xianglin, a colorectal surgeon at Kaohsiung Medical University Hospital, highlighting a clinical pain point. This is especially true for tiny or flat polyps, which are nearly invisible to the naked eye. "Using an AI Agent as a second layer of safety can improve the detection rate for such lesions."

To this end, the hospital assigned six physicians to assist in cross-labeling and validating colonoscopy screening data, transforming it into assets. Foxconn then helped train the AI model. It took approximately one year for the colonoscopy AI Agent to take initial shape.

We were granted exclusive, first-time access to the laboratory where this AI Agent was born, which resembles a consultation room. A bed holds a mannequin and endoscopy equipment, while a computing host jointly developed by Foxconn and Congtai Technology, powered by NVIDIA IGX Thor, is set up beside it, allowing the team to simulate real doctor operation scenarios.

During screening, the system can transmit images to a large screen with microsecond-level ultra-low latency. If a polyp is detected, it issues an instant audio alert, judges its size and pathological classification, marks the lesion location, and creates a 3D digital twin of the intestine.

Additionally, it can calculate and alert about withdrawal speed, collect quality indicators like bowel preparation cleanliness, automatically generate pathology reports, and, paired with a chatbot, assist in shortening the learning curve for junior doctors.

The development of the colonoscopy AI Agent relies on two key contributors: Congtai Technology, responsible for image selection and transmission, and Yuanyou Industrial, which controls the clinical distribution channels.

"We started planning even before the NVIDIA IGX T7000 was ready," recalled Lin Hongpei, General Manager of Congtai Technology. As a key partner for NVIDIA in promoting ultra-low latency transmission technology, Congtai became one of the earliest companies in Taiwan to access the IGX platform through its collaboration with Foxconn, "giving us about a six-month lead over other industry peers."

Recalling the platform migration period, Lin Hongpei chuckled, mentioning that whenever issues arose to ensure perfect hardware-software integration, engineers from both sides had to carry the scarce IGX Thor test hosts—only a handful in Taiwan—back and forth on Taipei's MRT Bannan Line. Within two to three months, they successfully migrated Foxconn's system to the NVIDIA platform.

Pan Zhenrong, General Manager of Yuanyou Industrial, the distributor for the globally dominant endoscope brand Olympus, emphasized the "value of localized AI." He noted that while Olympus, originating from Japan, has its own developed colonoscopy AI辅助 system, such foreign-developed AI systems can face adaptation issues in Taiwan. For instance, models in Taiwan need enhanced training to specifically recognize intestinal debris, flat polyps, and precancerous lesions.

Currently, in retrospective studies, this system has achieved an 82% sensitivity rate for polyp detection. Accuracy continues to climb as more data is fed into the model.

Although this AI Agent is still in the model training phase and not yet in routine clinical use, Dr. Cai Xianglin revealed that the next stage involves enabling the AI Agent to directly determine polyp types, predict the likelihood of precancerous lesions during the examination, and even assist doctors in deciding whether to perform an endoscopic resection on the spot. "We don't just want it to be a detection AI; we want it to evolve into a decision-support AI."

Returning to the nursing station scenario, Foxconn's most mature AI medical robot to date remains the Nurabot, which has completed validation at Taichung Veterans General Hospital. It reliably performs specimen delivery, medication transport, and patient education guidance, helping nurses reduce their movement time by approximately 30%.

Jiang Zhixiong emphasized that "Nurabot 2.0" has integrated VLA and NemoClaw architectures, enabling natural language communication and possessing capabilities like spatial understanding and task reasoning. In the future, nurses will no longer need to look down and tap on an app interface; they can free their hands and give instructions to the robot using natural language.

"Over the past two years, most AI existed on screens. Now, it's transitioning into the real world," said Jiang Zhixiong. While Physical AI is not a new term, its actual implementation in hospitals requires more rigorous validation and integration with existing Hospital Information Systems. Nurabot 2.0 represents Foxconn's latest attempt to advance Physical AI in hospitals, while other scenarios still face multiple hurdles, including varying hospital systems, incentives for National Health Insurance adoption, and product certifications like SaMD and TFDA.

As medical AI moves towards multi-faceted integration, it increasingly relies on collaboration across the entire ecosystem. "This time, Foxconn is not just a manufacturer of AI servers; we aim to further become an integrator of smart healthcare infrastructure," stressed Jiang Zhixiong.

This vision is gradually taking shape in the COMPUTEX exhibition area, where each scenario—from the chemotherapy room and operating theater to the consultation room and nursing station—serves as Foxconn's annotation on the era of Physical AI.

Facing a future of healthcare staff shortages, the entry of Physical AI into hospitals is inevitable. How to shift from a single-hardware mindset to multi-point integration continues to test the resolve and wisdom of this world's largest contract manufacturer.