Menu

Scenario robots — campus inspection rover & industrial autonomous material handling AMR

Two scenario robots from perception to deployment: inspection rover (LSTM, YOLO, PyQt/Streamlit/mobile + DeepSeek) and industrial AMR (PID, MobileNet).

Computer VisionRobotics

Nov 2024——Dec 2025

Background

This project targets deployable mobile robots in campus O&M and industrial logistics under the college’s innovation practice program, threading perception–decision–actuation–human–machine coordination into deliverable integrated systems. The campus track stresses time-series modeling of multi-source environmental data, vision-based safety, and multi-client operations under open environments, pedestrian flow, and weather variability; the industrial track stresses cycle time, positioning accuracy, and reliable material recognition, with motion control and lightweight vision co-deployed on embedded compute. Milestones are driven by university smart-car competitions and platforms such as the iCAN International Invention & Innovation Contest—covering requirements analysis, prototype iteration, and field integration, with algorithm validation, embedded implementation, and demo readiness under one timeline; on the handling side, a literature review synthesizes machine vision for mobile manipulators to support later papers and patents, forming a loop of competition-driven iteration plus academic consolidation.

  1. For drifting campus environments and slow patrol response, led a bidirectional LSTM + multi-head attention forecaster holding 1.8–4.2% error and cutting seasonal error from 31% to 6.5%; built a PyQt5 host with cloud connectivity, pedestrian detection (99.3%), and 0.5 s braking within 1.5 m.
  2. For strict logistics recognition and tracking tact, deployed PID line-following and embedded MobileNet on the autonomous transport platform with full vehicle bring-up, calibration, and smart-car contest rehearsals at showcase-ready metrics.
  3. For material-handling perception bottlenecks, led a full-stack framework and a ~100-paper survey comparing classical geometry with deep learning, analyzing SLAM, SDF, and visuo-tactile fusion barriers on embedded platforms, supporting the core-journal Robot Technique and Application survey.
  4. As technical lead, coordinated system tests, deployment docs, and issue closure on both robots, driving prototype iterations and field tuning via innovation contests including Internet+ and iCAN.

Research outputs

Stack

Python, PyQt, Streamlit, MobileNet, YOLO