I'm trying to implement a collision detection system to help solve constrained non-linear optimization problems in robotics as my master thesis.
Basically, what I need out of the system is the penetration depth and contact normal between each pair of potentially colliding bodies. I was thinking of using the GJK algorithm as a basis to implement the narrow phase of the system on GPUs using CUDA/OpenCL but wasn't sure if this line of thinking would get me very far due to the inherent lack of ‘numerical robustness’ of GJK using 32 bit floating point representations (the …