RL-ACD: Reinforcement Learning-based Approximate Convex Decomposition

Approximate Convex Decomposition (ACD) aims to approximate complex 3D shapes with convex components, which is widely applied to create compact collision representations for real-time applications, including VR/AR, interactive games, and robotic simulations. Efficiency and optimality are critical for ACD algorithms in approximating large-scale, complex 3D shapes, enabling high-quality decompositions with minimal components. Unfortunately, existing methods either employ sub-optimal greedy strategies or rely on computationally intensive multi-step searches. In this work, we propose RL-ACD, a data-driven, reinforcement learning-based approach for efficient and near-optimal convex shape decomposition. We formulate ACD as a Markov Decision Process (MDP), where cutting planes are iteratively applied based on the current stage's mesh fragments rather than the entire fine-grained mesh, leading to a novel, efficient geometric encoding. To train near-optimal policies for ACD, we propose a novel dual-state Bellman loss and analyze its convergence using a Q-learning algorithm. Comprehensive evaluations across diverse datasets validate the efficiency and accuracy of RL-ACD for convex decomposition tasks. Our method outperforms the multi-step tree search by 15× in terms of computational speed, while reducing the number of resulting components by 16% compared to the current state-of-the-art greedy algorithms, significantly narrowing the sub-optimality gap and enhancing downstream task performance.