動的環境におけるロバストなロボット把持のための自己教師あり学習
Self-Supervised Learning For Robust Robotic Grasping In Dynamic Environment
RGBDセンサと固有受容感覚データに自己教師あり学習を適用し、動的環境での把持戦略をリアルタイムに学習・改善する手法を提案。既存手法より把持成功率が15%向上した。
著者: Ankit Shaw
分類: cs.RO
原文アブストラクト
Some of the threats in the dynamic environment include the unpredictability of the motion of objects and interferences to the robotic grasp. In such conditions the traditional supervised and reinforcement learning approaches are ill suited because they rely on a large amount of labelled data and a predefined reward signal. More specifically in this paper we introduce an important and promising framework known as self supervised learning (SSL) whose goal is to apply to the RGBD sensor and proprioceptive data from robot hands in order to allow robots to learn and improve their grasping strategies in real time. The invariant SSL framework overcomes the deficiencies of the fixed labelling by adapting the SSL system to changes in the objects behavior and improving performance in dynamic situations. The above proposed method was tested through various simulations and real world trials, with the series obtaining enhanced grasp success rates of 15% over other existing methods, especially under dynamic scenarios. Also, having tested for adaptation times, it was confirmed that the system could adapt faster, thus applicable for use in the real world, such as in industrial automation and service robotics. In future work, the proposed approach will be expanded to more complex tasks, such as multi object manipulation and functions in the context of cluttered environments, in order to apply the proposed methodology to a broader range of robotic tasks.
