Intelligent Controller Integration for Heterogeneous Unmanned Systems

Project background

This work formed part of Science and Technology Innovation 2030 — New Generation Artificial Intelligence, a national major project in China.

Item	Description
Research task	Integration and validation of high-end intelligent controllers for cross-domain heterogeneous autonomous unmanned systems
Parent project	Open, general-purpose high-end intelligent controllers for autonomous unmanned systems
Task period	June 2022 – May 2025
Program management	High Technology Research and Development Center, Ministry of Science and Technology, China
Demonstration platform	Multi-rotor UAV with a robotic manipulator

The dates above describe the research task’s execution period. My involvement was through research work at Tsinghua University, beginning in 2024.

The project investigated adaptable intelligent controllers and their hardware/software integration across heterogeneous unmanned platforms. The UAV–manipulator demonstration brought together perception, mapping, and manipulation for field validation.

My engineering work

My work included deploying SLAM algorithms on onboard computing platforms, integrating robot communication, and supporting camera–arm calibration and object-grasping demonstrations. This involved visual and LiDAR mapping workflows, including ORB-SLAM and FAST-LIO, and integration of a depth camera with the robotic manipulator.

Validation overview

The presentation documents three areas of platform validation:

Object recognition: A UAV-view dataset containing 2,200 images across five object categories for an explosive-ordnance-disposal application, with a YOLOv5 demonstration on the controller.
Manipulation: Repeated object-grasping demonstrations using the integrated UAV–manipulator platform. The presentation reports a validation procedure of ten successful grasping cycles for each of four object categories, with object positions changed between attempts.
Outdoor mapping: A FAST-LIO field trial covering an area exceeding 1,000 m², illustrated with outdoor flight footage and reconstructed point clouds.

The task’s assessment requirements included a target detection rate of at least 95%, at least three supported functions, and SLAM coverage of at least 1,000 m², to be assessed through third-party testing. These are assessment requirements, rather than a claim that the clips below constitute a full acceptance report.