A mapless depth-image query speeds reactive drone collision checks

When a drone's planner has only milliseconds to react, should it first build a map to make collision checks faster? In short-horizon planning, the map's preparation time may outweigh its query savings, so the relevant question is total time from a new depth measurement to a collision answer.

The authors replace per-update map building with GPU distance queries over buffered depth images, aiming to answer a reactive planner's collision checks sooner.

On recorded forest depth sequences, the method answered 400 queries in 3.3–6.2 milliseconds on an Orin NX, with lower total latency than each measured voxel grid. This supports the idea that avoiding map preparation can help when a reactive planner makes relatively few checks per sensor update.

In branch simulations that charged measured computation costs, the evaluator recorded fewer collisions with the proposed method than with FIESTA or a KD-tree. The rates were percentages of runs averaged over five speeds, supporting the method's value in this tested setup.

The collision-rate comparisons come from simulated forest runs with measured Orin NX computation charged; they do not establish rates for other environments or real flights.

The real-world demonstration covered ten flights at two forest sites, and the method assumes the surroundings stay static over its buffered planning horizon.