Learning
Notes on automation, robotics, papers, experiments, and engineering tools.
This timeline records when the learning or project work actually took place, rather than when the website entry was written. Full technical notes are maintained in Chinese; the English summaries below provide the same chronological overview.
Learning Timeline
July 2025 — STM32 peripherals and communication
I connected the basic roles of GPIO, external interrupts, UART, and I²C into a single microcontroller learning path. The main lesson was to relate every software configuration to its electrical behavior and to select polling, interrupts, or communication peripherals according to the system requirement.
September 2025 — Modern optimization algorithms
I compared simulated annealing, genetic algorithms, and ant colony optimization through their search mechanisms, parameters, strengths, and failure modes. This study shifted my attention from simply choosing an algorithm name to defining the representation, objective function, constraints, and repeatable evaluation protocol.
April 2026 — Intelligent rescue vehicle
For a university engineering competition, our team developed an autonomous rescue vehicle and received Second Prize. The system evolved from a K210-based vision prototype to a final OpenMV4 H7 Plus and STM32F407 architecture, integrating color perception, a task state machine, mecanum-wheel motion, encoder feedback, and PID control.
April–July 2026 — Reproducible engineering toolchain
I organized a workflow covering isolated AnyLabeling environments, human verification of automatic annotations, scoped AI-assisted coding, and a local RGB-D dataset manager. The key principles were immutable raw data, explicit directories, fixed class mappings, batch manifests, and verifiable commands.
May 2026 — RealSense D435i RGB-D acquisition
I studied how to collect blueberry RGB-D data that could support maturity recognition, instance segmentation, and 3D localization. The workflow included stream alignment, camera intrinsics, scene metadata, depth-quality checks, structured storage, and controlled coverage of distance, viewpoint, illumination, maturity, and occlusion.
July 2026 — Blueberry instance segmentation and 3D localization
I reframed the perception task from detecting blueberries to estimating per-fruit masks, maturity, robust depth, and 3D position for robotic harvesting. The current work separates proposed research directions from completed results and emphasizes baselines, layered evaluation, ground-truth measurement, and hand–eye calibration.
July 2026 — Reproducible YOLO segmentation and error analysis
I standardized annotation conversion, dataset splits, experiment metadata, training, validation, and failure analysis. In one YOLO11s-seg test, most errors on semi-ripe blueberries came from confusion with the two neighboring maturity classes rather than missed detections, motivating clearer labels and condition-aware evaluation.