This paper splits a drone swarm into two roles and trains both with a role-based variant of MADDPG. Pursuers track several evaders that are learning to escape and are 30% faster. Scouts explore, rewarded for shrinking the largest Voronoi cell in the team so that they spread over the arena. Agents pass target locations to neighbours over a communication network, and policies are trained in the Multi-Agent Particle Environment before deployment on Crazyflie drones.
Findings
- Rewards: +20 to a pursuer that reaches a target, −10 for collisions, plus distance shaping and a soft boundary penalty for a 6 m × 6 m arena.
- With two targets and no scouts, the mean minimum distance to targets fell from 0.63 m with two pursuers to 0.46 m with five and barely moved with six, so the authors fixed five pursuers (sensing range 0.5 m).
- Five scouts gave about 35% average sensory coverage of the arena.
- Adding scouts left tracking roughly unchanged: mean minimum distance to targets was 0.46 m for plain MADDPG and 0.47 m for the role-based version.
- Real flights used six Crazyflie 2.1 drones (3 scouts, 2 pursuers, 1 evader) with eight UWB anchors and three cones as obstacles; the average pursuer distance to the target from UWB logs was 0.50 m.
Limits
- 2D motion and a small arena. Coverage of about 35% is modest, and the authors say coverage optimisation is out of scope.
- Target information propagates over a communication network that is assumed to work; dropped or false messages are not tested.
- Reward values were set empirically.
Relevance
The learned aerial swarm behaviour here runs on Crazyflies and trusts every teammate’s report. Pursuers act on target locations relayed by neighbours, so one robot reporting a false target would pull the team; how much is an open question for resilient consensus style filtering. The follow-up kouzehgar-2025-networked-pursuit-evasion drops communication at deployment altogether.
Links
- arXiv · video
- Subfield: Multi-robot systems, Robot learning, Field robotics