Robotic collision detection accelerators suffer from Orientation Aliasing and Granularity Mismatch with multi-sphere models. We propose S-COPU, a sphere-level speculative scheduling architecture using a “generate-then-predict” flow. S-COPU aligns prediction granularity with geometric representation, reducing geometric queries by 25% and end-to-end latency by 14.9% relative toCOPUin high-clutter environments without increasing memory requirements.
Lan et al. (Thu,) studied this question.