eNeural Technologies showcases eSL-Craft, a self-learning visual AI pipeline designed to keep edge models continuously up to date without manual retraining. Built on Advantech’s WA AIoT software platform and H-box edge systems, the service automates data collection, model retraining and deployment so inspection, logistics and robotics workloads can adapt to new conditions directly at the edge pipeline. https://www.eneural.ai/
At the core of the demo is eSL-Craft, a patented “adaptive self-learning” system that can refine object detection and segmentation models on real production data with minimal human labeling. It is part of eNeural’s broader AI-Craft toolchain, which combines automated annotation, model architecture optimization and quantization to shrink edge models while preserving accuracy and cutting AI time-to-market by as much as six times, particularly on NPU and embedded vision platforms using compact convolutional networks and mixed-precision training.
In this SPS 2025 demo at the Advantech booth in Nuremberg, a camera plus H-box edge GPU tracks pallets in a warehouse scenario and detects “corner cases” when new objects or packaging types appear. Instead of exporting raw video to the cloud, the system logs only the relevant samples, schedules on-device retraining on a zonal master node and pushes updated models back to the local inference devices, creating a closed loop between perception, data selection and continuous model refinement inside the factory line.
The architecture follows an edge-first and zone-based pattern: multiple cameras and NPU-powered clients handle real-time inference, while a GPU-equipped master node in each zone performs periodic retraining and then synchronizes only model weights with a central server. This resembles federated continual learning in industrial environments, keeping sensitive imagery on premises while still aggregating model improvements globally for robust multi-site deployment.
Looking ahead, the team discusses extending the same pipeline to mobile and humanoid robots that observe people and environments during the day, then retrain locally while charging at night to learn new faces, layouts and behaviors without exposing raw personal data. Today the heavy training runs on GPUs, with NPUs dedicated to low-latency inference, but the roadmap clearly targets broader use across smart factories, warehouses and service robotics, delivered as a combined Advantech hardware plus eNeural self-learning AI subscription for industrial customers.
