Detecting Mosquito Larvae Risks at Tropical Construction Sites with AI Robots : Clobot’s AI PoC Story
- SmileShark Team
- 23 hours ago
- 5 min read
[ 💡Key Highlights ]
1. Built an AWS-based AI proof of concept (PoC) environment in collaboration with SmileShark
2. Validated a prototype that automatically identifies mosquito larvae–prone environments using a VLM and RAG approach
3. Significantly reduced model comparison and testing time using Amazon Bedrock and Amazon SageMaker
4. Confirmed the potential to automate compliance checks against pest control guidelines using a knowledge-based approach
Clobot Co., Ltd.
회 사 명 Clobot Co., Ltd.
사업분야 Robot services and robot solutions
설 립 일 May 8, 2017
Clobot provides robot services across multiple industries, including guidance, delivery, cleaning and disinfection, safety and patrol, logistics, and manufacturing. These services are built on its heterogeneous robot autonomous navigation solution, Chameleon, and its robot management and control platform, CROMS.
Through a global partner network, Clobot sources and supplies robot hardware tailored to customer needs, including service robots, logistics robots, and industrial robots. The company integrates core technologies such as object recognition and AI perception modules for depth (distance) and pose estimation to enhance autonomous driving and motion control. Clobot continues to expand its solution portfolio with AI-based remote monitoring and guidance services, as well as human–robot interaction (HRI), optimized for the requirements of each industry.
Building Robots That See, Understand, and Respond : Clobot Robot AI Technology Team
Q. Please introduce your team.
Our Robot AI Engineering team primarily develops vision-based AI models to recognize objects and estimate attributes such as pose and distance. These models are integrated into autonomous robot solutions and used to improve obstacle avoidance and advanced robot motion generation.
Based on these capabilities, we train and deploy AI models tailored to real-world scenarios, including anomaly detection for patrol robots and human–robot interaction, such as social HRI. More recently, we have also been working on projects that incorporate Vision-Language Models (VLMs).
Q. What led you to pursue an AI PoC with SmileShark?
Our review of an AI PoC began in earnest after AWS Summit Seoul 2025. Following the event, our interest grew in the capabilities provided by Amazon Bedrock and other managed AI services. Through a proposal from AWS, discussions around a PoC with SmileShark began.
At the same time, Clobot was advancing a project to deploy disinfection robots at tropical construction sites to proactively address pest-related risks to worker safety. In particular, mosquito-borne infection risks emerged as a critical concern, creating the need for a technical approach that could assess environments with a high likelihood of mosquito larvae formation in advance.
Limitations of Existing Approaches
Q. What was the first challenge you wanted to validate in this PoC?
Tropical construction sites experience frequent heavy rainfall due to squall-like weather patterns. As a result, conditions favorable to mosquito larvae can form in many places, including standing water on the ground, uncovered trash bins, and damaged barricades.
With conventional vision-based object detection or segmentation models, it was difficult to assess such diverse situations using a single, unified criterion. We therefore aimed to verify whether combining a Vision-Language Model (VLM) with knowledge-based information would allow robots to move beyond simple object recognition and make holistic judgments about whether an environment is prone to mosquito larvae.
A key validation point was whether the VLM could not only determine the risk level but also present clear reasoning for its judgment. To do this, we used pest control guides and training materials—previously used for educating disinfection managers—as a knowledge base, and analyzed real-world images captured by on-site cameras.
Q. Were there additional goals beyond technical validation?
In parallel with validating the technology itself, we wanted to assess the time, cost, and performance required to implement the solution in an AWS-based environment. Another important criterion was how easily such an architecture could be built within the AWS Management Console, with a relatively low implementation barrier.
An Agile AI PoC Project
Q. How was the PoC conducted?
This PoC was carried out using an agile approach. We first defined high-level scenarios and designed the overall system architecture based on them.
We then worked through short sprint cycles to quickly learn and test services such as Amazon Bedrock, AWS Lambda, Amazon OpenSearch Service, and Amazon SageMaker. Each functional component was implemented and validated incrementally.
Once the core functionality reached a certain level of maturity, we implemented workflows using LangChain and LangGraph, and tested the end-to-end flow through a Streamlit-based user interface.
A Rapid Experimentation Environment Powered by Bedrock, RAG, and SageMaker
Q. What were the key outcomes of this PoC?
Through this PoC, we established a foundation for rapidly experimenting with and comparing multiple AI models in environments close to real-world conditions. By leveraging the various foundation models pre-integrated into Amazon Bedrock, we were able to easily compare and test model performance.
In addition, the fully managed RAG architecture enabled us to quickly test chatbot scenarios without building separate infrastructure. As a result, we significantly reduced the time required for initial experimentation.
We also confirmed that, in a cloud environment where costs scale based on actual usage, it is possible to repeatedly test a wide range of models without deploying high-specification training servers.
Q. What new insights did you gain?
First, we clearly validated the scalability of knowledge-based VLMs combined with RAG. This approach can be used to automatically classify compliance with pest control guidelines and can be extended to generate reports identifying suspicious areas, high-risk zones, and required remediation actions.
We also found the RAG-based architecture to perform better than expected. This result demonstrated strong potential for application across various domains.
In addition, we confirmed that Amazon SageMaker JumpStart enables relatively fast fine-tuning of pre-trained models and endpoint deployment. While we already operate an internal MLOps framework, we found that Amazon SageMaker Pipelines–based experimentation also has clear potential for practical use.
SmileShark as a Trusted Partner
Q. What stood out most about working with SmileShark?
What impressed us most was the fast and accurate support system. Technical and general inquiries were clearly separated, allowing us to receive prompt responses from the appropriate experts.
Even for questions that required more time, SmileShark provided detailed explanations and thoughtful feedback, which helped build a strong sense of trust throughout the collaboration.
Advice for Teams Starting Their First PoC
“ It’s more important to focus on the problem you need to solve than on the technology itself. ” - Ji-hyun Kwon, Team Lead, Robot AI Technology Team
“Because PoC timelines are short, it’s critical to start with clearly defined goals.” - Hyun-woo Lee, Senior Researcher, Robot AI Technology Team
The Next Step After the PoC
Q. Do you plan to apply AI technologies to real services or operations?
To apply AI technologies to real services or operations, several key elements must be in place. These include an end-to-end automated flow that connects sensor data collection and transmission, AI-based analysis and decision-making, monitoring, and robot behavior control, as well as an operational framework that can run this flow reliably.
Evaluation processes must also be incorporated, along with safety standards and recovery mechanisms tailored to different situations.
As we move toward real-world deployment, we plan to gradually complement domain-specific sensor integrations, evaluation and logging in operational environments, and the establishment of safety standards, ultimately connecting the PoC to actual services.
As a next step, we look forward to working again with SmileShark on AWS architecture design, security, and operational scalability.
SmileShark continuously supports companies like Clobot Inc. so they can focus on their business.
