Learned cut quality for approximate convex decomposition

Investigate the quality of learned cutting-plane proposals for approximate convex decomposition, particularly to determine whether trained policies can improve upon the hand-designed search heuristics used by CuACD.

Background

CuACD uses hand-designed search heuristics inherited from the CoACD lineage rather than learned proposals. The paper acknowledges that some resulting cuts are suboptimal compared with those that could be selected by a trained policy, leaving the quality of learned cut proposals unresolved. This problem concerns evaluating or developing learned policies that can improve the geometric quality of the decomposition while retaining the efficiency advantages of the fully GPU-resident CuACD pipeline.

References

CuACD retains the hand-designed search heuristics of its CoACD lineage and does not use learned proposals; some cuts are suboptimal relative to a trained policy, and we leave learned cut quality to future work.

— CuACD: A Fully GPU-Resident Approximate Convex Decomposition  (2609.28731 - Shi et al., 23 Sep 2026) in Section Conclusion and Discussion, paragraph “Limitations and future work”