Cuboid geometry for end-effector attachments?

[onurcalisir27]

Hi,
I'm exploring VAMP for pick-and-place applications and have been impressed with the performance numbers. I noticed that while the environment supports cuboid obstacles (with sphere-cuboid collision checking), the attachment API currently only supports sphere collections via add_sphere().

Is there interest in supporting cuboid geometry for attachments? For manipulation tasks involving box-shaped objects, sphere approximations can be quite conservative and may cause unnecessary false positives in tight spaces.

I understand this would require implementing cuboid-cuboid collision checking (environment cuboids vs. attached cuboid), which is more complex than sphere-cuboid. I'm curious whether:

1. This has been considered and deprioritized for a reason (e.g., branching costs breaking SIMD efficiency)?
2. It's something that could be contributed, and if so, any guidance on the approach?

Separately, I noticed the roadmap mentions mesh collision but not distance queries. Is clearance-aware planning (minimum distance to obstacles rather than binary collision) something being considered for future versions?

Thanks for the excellent work on VAMP.

[zkingston]

Thanks for your interest! There is interest in supporting additional geometry in the attachment and on the robot. This would require implementation of additional collision kernels, as currently only sphere/primitives collisions are implemented. There isn't anything theoretical standing in the way of making this happen, and we plan to add these at some point in the future. There is some difficulty in implementing cuboid-to-point cloud collisions on top of the CAPT data structure, which probably would remain unimplemented for the foreseeable future.

As for clearance, see #89 for some progress towards this support.