Reading this primarily to see what ICRA papers are like
- The basic problem they work on is NP-Hard, so they are excused in using whatever method they like
- Assumption is each vehicle knows initial trajectories of other vehicles in terms of waypoints
- MC method (not sure which) is then used to estimate uncertainty of actual trajectories, due to atmospheric conditions, model of each UAV, and sensors
- A collision-free path planning algorithm based on GAs is proposed, which plans in terms of waypoints
- The planning alg takes into account the uncertainty estimates which resulted from variance from MC
- The approach is non-cooperative, which means when a future collision is predicted, a vehicle changes direction independently without consulting other vehicles
- Considered scenario is navigation in the 2D plane, velocity is not permitted to change
- Collisions are estimated by gridding the space and checking closeness in terms of number of grid cells separating the two
- Fitness function of GA factors in: length of traj, deviation of traj from initial plan, distance of trajectory from static obstacles, turning angle between consecutive segments, and whether a collision occurs
- It is extremely strange that they say the weight for deviation is the by far the larges. I would assume the weight related to collision is the most important
- There are some additional constraints: the trajectory is constrained to a maximum lengh, maximum amount of turns in path, and that collision is not allowed (I don’t understand why there is a related weight in the fitness function then)
- They say they get near optimal results from this – which seems possible, I don’t think its so difficult
- There is another set of criteria that is evaluated when trying to fly waypoints with MC; want low (statistical) deviation of path from original, distance from obstacles, and final location must be close to goal
- Empirical execution time is linear in number of vehicles, but constant in number of static obstacles
- I see why the paper got in now – they did it on a real UAV doing collision avoidance with simulated other UAVs. I think computation is offboard