Embodied Drones for City Maintenance

ProblemCleaning an irregular high-rise glass façade by hand takes about two weeks each time, with workers at height.

• Safety-assured UAV positioning, perception and control near glass façades
• Purpose-built cleaning drone designed with the PolyU Campus Facilities team
• Spun out as the CeresRobotics.ai start-up for building-exterior cleaning

ResultsSilver medal at the International Exhibition of Inventions of Geneva 2025; PCT, US and CN patent filings; covered by TVB and Ming Pao (May 2024) and UBeat (Nov 2025); student Innovation & Entrepreneurship Third Prize (2024).

ProblemNear a wall, GNSS is blocked, LiDAR sees only one plane and airflow pushes the drone toward the glass.

• IMU–GNSS–LiDAR–camera fusion tuned for façade-proximity flight
• Wall-proximity aerodynamic compensation in the controller
• Flight tests on PolyU buildings with the cleaning drone

ResultsSupports the CeresRobotics.ai cleaning-drone product.

ProblemLow-altitude operations need UAVs that position reliably and understand their tasks in dense urban airspace.

• High-accuracy UAV positioning for low-altitude operations
• Vision-language-action (VLA) models for intelligent UAV tasks
• Builds on TAS Lab drone platforms and the low-altitude-economy regulatory sandbox flights

ResultsLinks to the PolyU MSc in Low-altitude Economy and the PolyU-Wuxi drone product lines.

ProblemOffshore wind turbines are far from shore and GNSS-challenged near blades and towers; manual inspection is slow and risky.

• Autonomous UAV system for blade and tower inspection of offshore wind turbines
• Robust positioning and perception near large rotating structures
• Developed within the Guangdong–Hong Kong Joint Laboratory for Marine Infrastructure

ProblemSlopes, forests and civil structures block satellites, yet inspection drones need certified positions to fly safely.

• Rotary-wing UAV with camera, solid-state LiDAR, multi-frequency GNSS, IMU and onboard computing
• LiDAR-aided GNSS with NLOS mitigation and factor-graph pose estimation
• Deep-learning detection of landslides, rockfalls and building damage from images and point clouds

ResultsThe positioning work fed the cleaning-drone US patent application.

ProblemCity maps go out of date quickly; ground robots lose GNSS in street canyons and cannot see the big picture.

• Bandwidth-light UAV–UGV feature exchange, with UAV GNSS as the ground robot's reference
• Factor-graph fusion of LiDAR, camera, IMU and GNSS from air and ground
• A large vision model detects map changes for update

ProblemLearned controllers fly aggressively but give no safety guarantee; classical planners are safe but slow and conservative.

• Deep networks that plan and control directly on SE(3), position and attitude together
• Safety filters that keep learned commands inside certified bounds
• Validation on TAS Lab agile UAVs

ProblemDrones in cities need somewhere safe to land, recharge and be managed, without a human on site.

• UAV airports on smart street-light poles with wireless charging and AprilTag-based precision landing
• Custom power-distribution board for seamless switching between battery and external power
• GNSS, Vicon and AprilTag localisation fused by an extended Kalman filter; 4G MQTT communication

ResultsExperimental validation of AprilTag localisation, power management and wireless communication under dynamic conditions.

ProblemLarge or high structures cannot be printed by fixed gantries; a drone printer must hold millimetre accuracy while flying.

• Rotary-wing UAV with positioning, LiDAR, a lightweight printer and a nozzle manipulator
• Joint MPC of rotors and manipulator, with extrusion matched to print speed
• In-process quality monitoring by solid-state LiDAR scanning

ResultsThe platform links to the lab's UAV inspection and cleaning work.