This paper proposes a Bayes Net based Monte Carlo optimization for motion planning (BN-MCO). Primarily, we adjust the potential fields determined by goal and start constraints to progressively guide the sampled clusters toward the goal and start points. Then, we utilize the Gaussian mixed modal (GMM) to perform the Monte Carlo optimization, confronting these two non-convex potential fields. Moreover, KL divergence measures the bias between the true distribution determined by the fields and the proposed GMM, whose parameters are learned incrementally according to the manifold information of the bias. In this way, the Bayesian network consisting of sequential updated GMMs expands until the constraints are satisfied and the shortest path method can find a feasible path. Finally, we tune the key parameters and benchmark BN-MCO against the other 5 planners on LBR-iiwa in a bookshelf. The result shows the highest success rate and moderate solving efficiency of BN-MCO.

翻译：本文提出一种基于贝叶斯网络的蒙特卡罗优化运动规划方法（BN-MCO）。首先，我们通过调整由目标约束和起始约束确定的势场，逐步引导采样簇向目标点和起始点逼近。随后，利用高斯混合模型（GMM）执行蒙特卡罗优化，以应对这两个非凸势场。此外，采用KL散度度量由势场确定的真实分布与所提GMM之间的偏差，并根据偏差的流形信息逐步学习GMM参数。通过这种方式，由序贯更新的GMM组成的贝叶斯网络不断扩展，直至约束得到满足，从而最短路径算法可找到可行路径。最后，我们调整关键参数，并在书架场景中的LBR-iiwa机械臂上对BN-MCO与其他五种规划器进行基准测试。结果表明，BN-MCO具有最高的成功率和适中的求解效率。