Deyang Lin

Consumer Hardware Pest Control Startup (2024-present)

Mon, 01 Apr 2024 00:00:00 +0000

Searching for a New Direction

After realizing that the robotics studio was becoming a stable business rather than a startup that could scale quickly, I began looking for a new direction that was worth long-term commitment. In April 2024, I joined XbotPark Camp, where I met many founders working on hardware products.

The first direction I explored was companion robotics. The central question was how to build a robot that could deliver a more natural sense of companionship, instead of stopping at AI conversation or visual styling.

After roughly two months of user interviews, we still could not find a strong software-hardware entry point. Around that time, I met Enzo through an investor friend. He was looking for a technical co-founder, and I was looking for a direction with real market pull. We eventually focused on Pest Control, a traditional market with persistent demand and significant room for technical improvement.

Why Pest Control

The previous studio experience taught me that choosing the right market matters as much as building the product. A good consumer hardware direction should have a large market, long-term demand, a reasonable competitive landscape, and a clear reason for software and hardware to work together.

My initial screening criteria were:

A large market with a high ceiling and persistent demand.
Avoiding direct competition with major home-appliance companies and highly saturated outbound hardware categories.
Focusing on global markets, where consumer hardware can have healthier margins.
Choosing a traditional field that has not yet been deeply modernized by technology.
Creating stickiness through hardware, software, services, and potentially recurring value.

Pest Control matched these filters. It is traditional, widely needed, and still underserved in safety, effectiveness, cleanup, placement guidance, remote notification, and intelligent management.

Because the product has not officially launched yet, many details cannot be disclosed. At a high level, the work focuses on:

Intelligent IoT.
Safer and more humane handling.
Easier cleanup and post-processing.
Better placement guidance and continued usage.

Starting from Field Research

At the beginning, we did not raise money or scale immediately. Two of us started from the problem itself. This stage felt closer to research: the answer was not clear upfront, so we had to repeatedly validate assumptions in real environments.

We started from a small workspace in Hangzhou and spent most of our time on field research and prototype testing.

The first small office in Hangzhou.

We also worked closely with a pest control service company, following technicians into restaurants, kitchens, homes, and other real environments. At that stage, we were still deciding whether the first product should target consumers or service companies, so understanding the real workflow was the priority.

Field research with service technicians.

Studying existing tools and their limitations.

One of the early field-testing approaches.

Early Sensing and Test Platforms

One of the first insights was that the product needed a safer and more controllable way to detect, trigger, notify, and support cleanup. Around that goal, we worked on mechanical structure, circuits, embedded systems, sensors, and data logging at the same time.

An early data-collection device for video, sound, vibration, and infrared data.

The first test circuit.

We built multiple experimental setups to test different structures, detection methods, and trigger strategies. The work was fragmented but important because it forced the product definition to come from reality rather than imagination.

Early form-factor experiments.

Early testing equipment and experimental space.

Prototypes and Form-Factor Iteration

Over several months, we designed and tested dozens of different forms. The hard part was not a single technical component. The real challenge was knowing whether a solution would work in messy real-world environments and understanding the actual user decision process.

Early structure, circuit, and trigger experiments.

Building a dedicated test space.

B2B Validation and Lessons

One early direction was a long-form B2B device. It helped validate stronger on-site capability and remote notification, while also exposing practical issues such as safety detection, maintenance, water, oil, installation, and after-sales support.

Early B2B device testing.

The key lesson was that a strong device does not automatically mean the market will adopt it. B2B purchasing decisions, sales capability, existing business models, equipment cost, and value distribution all affect adoption.

That experience made it clear that hardware products cannot be evaluated only through engineering performance. The product must also fit the user’s incentives and decision logic.

Moving from B2B Capability to Consumer Product

After recognizing the limits of the B2B path, we moved part of the validated capability into a consumer product. Consumer users purchase and use the product themselves, so their needs for effectiveness, ease of use, cleanup, notification, and placement guidance are more direct.

We then moved through multiple rounds of consumer prototypes, iterating on size, sensors, circuits, cost, interaction, and safety strategy.

Around July 2025, we finished a prototype close to the final product direction. The team then returned to Shenzhen to work with industrial design partners, mold factories, and PCB partners with mass-production experience. The team gradually expanded from two people to five.

In this phase, I was responsible for product definition, prototype design, mass-production introduction, global multi-region IoT hardware and software architecture, app design, and building an AI-based internal development workflow for faster software-hardware iteration.

Continuing Product-Line Exploration

While pushing the consumer product toward production, we continued exploring other Pest Control needs, including additional rodent-related scenarios and mosquito-control products.

Takeaway

The most valuable part of this experience was not a single prototype or technology module. It was going through the full loop of direction selection, user research, field validation, hardware and software prototyping, B2B testing, consumer-product transition, mass-production preparation, and IoT system design.

It also made me more aware that hardware startup work does not end when a good device is built. The real challenge is repeatedly calibrating technology, product, market, supply chain, and business model until the team finds a direction that users genuinely need and that the team can keep delivering.

Robotics Studio (2022-present)

Fri, 04 May 2018 08:58:25 +0000

Robotics Studio (2022-present)

What I have got in the past

Since June 2022, I have co-founded a robot studio with friends in Guangzhou. The studio focuses on designing and teaching robotics solutions for secondary schools. In this studio, my primary responsibilities include the design, implementation, and teaching of comprehensive robot solutions (including structural design, with a personal focus on software design). Since 2023, the robots designed by the studio have won championships in competitions in cities such as Guangzhou, Foshan, Zhuhai, and Dongguan. We have delivered over 150 units to date.

The robots designed by the studio in 2023.

During the robot design process, we have integrated more advanced algorithms to maintain a strong competitive edge in the market with the same hardware. In the latest chassis design that I independently developed, I innovatively designed a bus motor control board based on I2C. This allows us to control a large number of motors, with support for up to 256 encoder motors, meeting the requirements of various robots.

Motor driver board design 1

Motor driver board design 2

Motor driver board design 3

Additionally, I designed a standardized communication protocol based on competition needs and integrated more advanced control algorithms into the motor drivers. For example, I used cascaded PID control and implemented trapezoidal velocity planning for precise motor control.

Furthermore, I developed an innovative Matlab-based upper computer that enables automatic motor identification, approximating the auto-tuning of velocity control PID. This greatly accelerated the overall development progress.

The implementation of excellent algorithms has resulted in outstanding motion performance for the chassis I designed. For instance, it achieves precise omnidirectional reciprocating motion within 2 meters with a cumulative error of less than 1 cm. This significantly accelerates the development of competition robots and reduces the difficulty of use for end-users. Moreover, the hardware cost of the solution remains comparable to conventional solutions in the market, greatly enhancing the competitiveness of our designed robots.

Furthermore, I have incorporated cutting-edge technology into robotics competitions for primary and secondary schools. I have developed a fully automatic image calibration system based on SAM, which enables automatic calibration of target patterns. This significantly improves the accuracy of recognition, addressing the issue of insufficient calibration images affecting recognition accuracy.

Pose estimation and point cloud perception (undergraduate thesis + Jilin University lab)

Fri, 04 May 2018 08:58:25 +0000

Pose estimation and point cloud perception (undergraduate thesis + Jilin University lab)

Pose estimation

My undergraduate thesis focused on the estimation of object poses. I input RGB images and the 3D models of the detected objects, mapping the 2D image pixels to the 3D point cloud of the model. Based on this, I utilized PnP (Perspective-n-Point) and RANSAC algorithms to regress the pose and achieve object pose recognition. Additionally, this approach incorporated a deep learning-based refinement algorithm to further improve the pose accuracy based on the initial pose estimation from PnP. Experimental results demonstrated that, compared to other related works, establishing the mapping relationship between the 2D plane and 3D space enabled higher accuracy in pose estimation.

Overall algorithm framework design

Creating a virtual dataset using randomly posed 3D models and COCO as the background

Dataset generation results

The principle of UV mapping

Mapping relationship between UV mapping and surface points (point cloud) of an object

Design of UV map generation network

Result of UV map generation

Comparison between UV map generation results and calibration images

Comparison between generated point cloud results and calibrated point cloud

The overall approach of using RANSAC+PnP for initial pose regression of objects

Design of a deep learning-based pose regression network

Pose recognition results

Point cloud perception

During my senior year, I primarily focused on point cloud perception in the laboratory at Jilin University. I accomplished the data format conversion from the Livox dataset to the KITTI dataset.

Data format conversion

Additionally, I successfully trained and implemented the PointPillars model using the Livox dataset for forward inference.

Forward inference and recognition Result1

Forward inference and recognition Result2

Through this experience, I gained knowledge of various methods for point cloud processing and became aware of their limitations. It also sparked my interest in exploring the fusion of point cloud and visual perception.

There is significant room for improvement in using purely point cloud-based methods for model recognition.

Wheeled-legged hybrid robot

Fri, 04 May 2018 08:58:25 +0000

Multi-degree-of-freedom wheeled-legged hybrid robot

“I have always been eager to build more complex robots. During my junior year, I joined the BIRL team at my university. I participated in the structural and circuit design of a leg-wheel hybrid robot and almost independently completed the entire project, including kinematic modeling and implementing the control code for the robot’s motion.

I conducted Matlab simulations based on the robot’s kinematics and solved forward and inverse kinematics using MoveIt in ROS.
I implemented proportional-velocity control for the three degrees of freedom of the robot’s motors using STM32 and CAN.
I achieved pose transformation for the robot’s motion in different planes.

Design and implementation of proportional-velocity (PV) control for a three-degree-of-freedom robot’s motors:

Testing the walking capabilities of the second-generation robot:

Real-world application scenario testing:

Projects in Robot Team

Fri, 04 May 2018 08:58:25 +0000

Several projects within the school robotics team

Since the beginning of my freshman year, I have joined the university’s robotics team and continuously improved my engineering and algorithmic skills through participating in numerous competitions.

Design of a grayscale line tracking board

We found that the robot is susceptible to external light interference (especially sunlight) when tracking outside of indoor environments. Therefore, I aim to construct a grayscale line tracking board that can resist external light disturbances.

Design of a dynamic line tracking board resistant to external sunlight disturbances:

Differential method is employed to dynamically turn off the board’s own light source, measure the intensity of external light, and then turn on the self-illumination to obtain the reflection intensity. By subtracting these values, the actual reflection intensity of the internal light source can be obtained, thus reducing the impact of external light.
Dynamic calibration of the white line/black background threshold is achieved using the Otsu algorithm.
This project has been granted a national utility model patent and is currently being used in various robotics competitions within the team.

The design of a grayscale line tracking board

The performance of the line tracking board under actual external sunlight disturbances.

Design and implementation of small-area robot localization

Implementation and application of robot localization algorithm based on encoder motors, gyroscope, and laser rangefinder. EKF (Extended Kalman Filter) is used to fuse data from encoders, gyroscope, and laser rangefinder. By fusing the data with encoders, the robot achieves centimeter-level indoor localization at a frequency of 50Hz, even with the laser sensor operating at only 10Hz.

Algorithm testing and visualization

The application of this localization algorithm in actual competitions

Jetson & YOLO

Design of a road vehicle counting algorithm based on Jetson Nano and YOLOv4-tiny.

Utilizing TensorRT to achieve recognition frame rates above 18fps on edge devices with power consumption below 15W;
Designing a data visualization interface based on the Node-RED framework".

Block-grasping robot

Fri, 04 May 2018 08:58:25 +0000

Project Introduction

In 2018, during the DJI High School Summer Camp, I served as the team captain and was primarily responsible for embedded development. Our main task during the camp was to design and manufacture a robot capable of grasping, storing, and stacking building blocks. In the project, my responsibilities included:

Coordinating and communicating with the structure, embedded systems, and algorithm teams to drive overall progress and resolve coordination issues.
In the embedded systems part, I completed the kinematic analysis and actual control of the two-link mechanical arm. This involved controlling the arm’s gripping mechanism based on STM32 and CAN to control the 3508 motors for the chassis and the 6002 gimbal motor. I also implemented material detection and control for material ejection.
Designing the communication protocol between the embedded system and the PC, which primarily focused on transmitting block corner position information for optimized grasping to the PC using serial communication and JSON protocol.

Overall robot solution design

Design and fabrication of the robot

Real-world testing of the robot