A Study Of Object Pose Estimation Based On Neural Network And Motion Planning Based On Sampling

Object pose estimation and motion planning are key technologies in many robotic tasks.On the one hand,previous pose estimation methods based on template matching or iterations are slow.Methods using neural networks always require a large model,which means it's not efficient in memory and computing.A light and accurate model is needed in real time embedding systems.On the other hand,robotic systems usually include various high dimensional motion planning problems.Therefore,algorithms that can deal with high dimensional problems and can be easily generalized to different scenes are preferred.For the pose estimation problem,we propose a method based on point clouds and neural networks.The model is small and accurate enough.We first sample local patches in the global point cloud.We propose a method for extracting key points from global or local point clouds and a method for pointwise feature calculation with the key points as reference points.This pointwise feature is invariant to object pose changes.We propose a hierarchical neural network based on PointNet,with the input of global and local unordered point set,to predict the object coordinates of key points.Both the pointwise feature invariant to pose changes and the point set data format in contrast to sparse grid data help reducing the complexity of input data.Therefore,the neural network can be lighter.In the end,according to the correspondence between the current and the predicted coordinates of key points of all the patches,we can estimate the pose by closed form solution.In the simulation experiments,about 2?3mm of the mean distance between closest point pairs is achieved.With only a few steps of ICP as post processing,we can further improve it to about 1mm.We show the benefits of the hierarchical structure by ablation test.In the end,we compare our method with others in the model size and the accuracy.For the motion planning problem,we choose a sampling based method.It replaces the random uniform sampler with the conditional variational auto-encoder.The new sampler can sample with greater probability near the optimal path by training in the dataset of successfully planned cases.The method can be easily generalized to different scenes by designing the conditional variable.Besides,we have proposed a sampler merging strategy and a low l2-disper sion sequential sampling strategy to further improve the algorithm.The algorithms are mainly tested based on OMPL.The CVAE based algorithm turns out to be more efficient than the original one.The two proposed strategies are also beneficial.