In this embedded systems study, YantriX investigated on-device computer vision inspection running on an ESP32-S3 microcontroller—evaluating INT8 neural inference within a compact low-power footprint for line-side quality checks.
Overview
Embedded vision study deploying quantized neural networks on a low-cost ESP32-S3 microcontroller to classify surface and assembly defects locally without PC hardware.
Discipline: Robotics & Automation
Project Type: Embedded AI (Microcontroller Vision + Quantization Study)
Application: Manufacturing Quality Inspection
YantriX Scope Ownership
Key specifications
Objectives
Challenge
Traditional line-side computer vision inspection relies on expensive industrial PCs, bulky gigabit cameras, and facility network drops. This study evaluated whether recent dual-core microcontrollers with vector instructions (ESP32-S3) can handle localized pass/fail surface inspection directly at the workstation within a low-power, self-contained footprint.
Approach
Trained a lightweight MobileNet-derivative architecture on custom industrial part images.
Applied full integer post-training quantization using TensorFlow Lite for Microcontrollers.
Utilized ESP-NN optimized vector assembly kernels on the Xtensa LX7 processor cores.
Designed a compact 40 x 40 x 25 mm snap-fit housing with adjustable mounting brackets.
Outcomes
Deliverables
Tools used
Impact
Conclusion
Thoughtful neural network quantization combined with embedded hardware co-design makes real-time computer vision feasible on constrained microcontrollers without requiring expensive PC infrastructure.
Working on a similar engineering problem?
Need compact, low-cost inspection or embedded perception hardware for your assembly workstations? Discuss embedded vision options with our engineers.
Tagged
Related case studies
Robotics · Vision-Language Models & Industrial AI
Computer vision study evaluating parameter-efficient fine-tuning (LoRA) on OpenCLIP for visual recognition across industrial mechanical fasteners.
Robotics · Vision Systems & Manipulation
Robotics engineering study demonstrating unstructured bin-picking using YOLOv11-Seg instance segmentation, 3D surface normal extraction, and collision-free grasp trajectory planning.
Continue exploring
Explore the engineering service and practical guides related to this work.
Discuss your on-device edge vision defect detection on esp32-s3 requirements, design challenges, or target specifications with our engineering team. NDA support is available where required.