M.Eng. Thesis, Tokyo Institute of Technology, 2022
Real-Time Optimization for Dynamic Multiple Target Allocation and Tracking with Heterogeneous Robotic Systems
Department of Systems and Control Engineering.
Part of Drone Coverage Control for Moving Target Tracking and Predictive Multi-Robot Task Allocation for Radiation Monitoring
Sending a team of robots after targets that move is really two problems, and this thesis takes on both. First, who should go where — allocation, when the tasks themselves are in motion. Second, how to not lose them — visual tracking, once a robot arrives. Both are posed as real-time optimization problems over heterogeneous UAV and UGV teams.
Part one — dynamic target allocation
Motivated by radiation monitoring over the La Africana mega-solar plant in Spain, where the “tasks” are moving cloud shadows. Allocating to a moving task is only correct if it accounts for where that task will be, so the method estimates task and robot position evolution, computes robot-to-task interception points, and allocates over a predictive receding horizon — commit the first step, then recompute.
The whole thing is relaxed into an equivalent linear program, which is what keeps it cheap enough to run online; energetic feasibility of each robot is folded into the same optimization. Validated in MATLAB and then in ROS/Gazebo on a model of the real plant, with 4 robots servicing 16 tasks.
Carried out with the University of Seville (Prof. J. M. Maestre and J. G. Martin), and published at ECC 2022 and in Solar Energy (2023). More on the project page.
Part two — dynamic target tracking
Prior work in the group used control barrier functions to hold a detected target in a drone’s field of view — which works while the target stays still. This part, motivated by collaboration with Fujitsu, asks what breaks when it moves, and finds the answer is a three-way coupling between altitude, detection model, and target speed.
Fly low and the detector is accurate but the field of view is too narrow to keep up; fly high and the target is too few pixels to detect. And the most accurate model is not the best tracker — at ~600 ms of latency the delay loses the target at any altitude.
Mapping that trade-off took 96 experiments — six TensorFlow detection models (30 ms to 620 ms), four altitudes (0.8–2.0 m), four target speeds (0–0.3 m/s) — flown on the Tokyo Tech Sky testbed with a Parrot Bebop 2 tracking a ground robot under OptiTrack. Those measurements turn the runtime choice into a small discrete optimization: estimate target velocity, then pick the altitude and detector that clear a performance threshold.
This line of work continued into SICE 2024 — see the project page.