Yeah, I was starting to realize it would be harder than I thought to modify the RVO...
But I think it would be a better idea to use a "local" RVO system instead of using Boids.
I actually already made a pathfinding system and an AI on a sphere for the game I'm working on, and I encountered several problems.
Here's the gameplay trailer, in which you can see some units moving:
To do the pathfinding system, I used an icosphere and a Dijkstra algorithm to link all the vertices together (no rocket science compared to navmeshes and such).
The AI was quite basic: it followed the path point by point (I actually rotated the agent around the center of the planet since it's a sphere instead of translating it), and I used a basic avoidance system that consisted in doing a SphereCast forward, and only steering left or right to dodge an obstacle or another agent (which led to agents stuck in walls or between other agents, so clearly not the perfect solution).
Also, I bought the Behaviour Designer Movement Pack because there are a lot of interesting components, and I thought I could avoid using the RVO controller thanks to some behaviours, but it turns out that they kind of relies on it too...
So that's why I think it would be a better approach on the issue to use a local plane for all the avoidance.
It could work to have, like you said, a plane for each agent (with the agent up vector as the normal of the plane), and project the positions of the surrounding agents/obstacles only this plane.
Do you have any advice or idea about how I could do that? Where could I implement this system? Which scripts to extend or modify?
Thanks again for your advice and your time.