ATLAS
From affine formation planning to real Crazyflie flights.
Submitted to IEEE Transactions on Automation Science and Engineering (TASE)
ATLAS learns to move and reshape a UAV formation using a reduced set of 13 affine actions. The planned trajectories are evaluated in simulation and executed by a team of five Crazyflies through leader–follower coordination.
- LearnPlan formation motion with an affine-action policy.
- SimulateReplay the trajectories with Crazyflie dynamics in MuJoCo.
- FlyExecute the position references on physical Crazyflies.
Sim-to-real demonstrations
Compare simulation replays with recorded flights in three matching scenarios. Each video has independent playback; the clips are not time-synchronized. Selected runs are shown here; the paper reports the full evaluation.
01 Sparse obstacles
A five-drone formation navigates around separated obstacles toward the target region.
02 Narrow passage
The formation traverses a passage bounded by square and hexagonal obstacles.
03 Shifted start and target
The trained policy is reused with shifted start and target positions and an additional obstacle.
This run uses slower trajectory playback and follower separation assistance.