Researchers beat me at my own game: a faster, cheaper point-cloud collision checker
We're not done with point clouds
Two years after publishing a point-cloud collision-checking data structure, the author finds that Chen and Yeh's multilevel voxel table (MVT) outperforms it on every benchmark. The MVT replaces nearest-neighbor search trees with a sparse voxel grid, cutting construction time and memory use. The author reimplements it in Rust, adds mutability, and discovers that the optimal voxel width is around 10–20 cm, not the largest robot sphere as originally suggested. End-to-end motion planning sees a respectable speedup.
If you wait long enough to solve a problem, someone else might just solve it for you.
- pj_mukh
The title implies that there are collision check systems that don’t use point clouds (or that robotics has moved onto something else)? Is that true or did the author just mean personally?
- ilikerashers
Why don’t they marry the visual categorisation of objects with a sparse predefined model to simplify?
Point clouds seem like a noisy layer to use and better as verification.
- fisian
Posts like these really highlight how much there to think about in robotics. Below the surface of what immediately comes to mind there is just an abundance of topics to address to get an actually well working robot.