Research On Active Learning Based Robotic Grasping Method

Robotic grasping is a basic function of intelligent robots and a challenging task.While a wide range of grasp pose detection methods have been proposed benefiting from the development of deep learning,a sufficient labeled dataset is significantly essential.Robots may move into different environments,and a new dataset is needed to re-train the model in order to maintain the performance when the environment changes.However,data annotation is a costly procedure.Active learning aims to mitigate the greedy need for massive labeled data by only selecting the most informative samples to be annotated,rather than annotating the whole dataset.This paper proposes a discriminative active learning strategy for robotic grasping.The strategy utilizes a shared encoder to derive latent features from both labeled data and unlabeled data.A discriminator is established to estimate the similarity between each unlabeled data and the labeled dataset.The lowest similarity stands for the most informative data sample.In consideration of the application in the real world,current grasp pose detection algorithms are not able to handle the balance between accuracy and real-time performance.This paper improves current grasp pose detection networks with depth-wise separable convolutions.The accuracy of the model is improved but the increase of parameters is limited to ensure realtime performance.Along with the proposed discriminative active learning strategy,an active learning framework robotic grasping is designed.The proposed discriminative active learning strategy and the grasp pose detection model are evaluated with two real-world grasp datasets.The proposed active learning strategy shows superior performance over other baseline strategies,especially when the amount of labeled data is relatively little.Considering annotation noise in real life,an experiment is performed on a noisy grasp dataset.The results demonstrate that the proposed active learning strategy is stable to noisy annotation.Compared to the other grasp pose detection models,the proposed depth-wise separable convolution based grasp pose detection model can achieve a better balance between the model accuracy and the real-time performance.Besides,real-world experiments are performed to show the superiority and the validity of the proposed grasp pose detection network and the framework of robotic grasping training and application.