KAIST Nonprehensile Objects

一种需要大量接触模式的非抓握操作系统过渡和利用环境接触成功将对象操纵到目标位置。我们的方法是基于深度强化学习不同于现有技术规划算法，不需要先验知识物体或环境的物理参数，例如作为摩擦系数或质心。计划时间减少到上的简单前馈预测时间神经网络。我们提出了一种计算结构，行动空间设计和课程学习方案促进了有效的探索和模拟到真实的转移。A non-prehensile manipulation system that relies on extensive contact patterns, leveraging environmental contacts and operational transitions to successfully manipulate objects to target positions. Our approach, based on deep reinforcement learning, differs from existing state-of-the-art planning algorithms in that it eliminates the need for prior knowledge of physical parameters of objects or the environment, such as friction coefficients or center of mass. The planning time is reduced to merely the duration required for a simple feedforward neural network to complete feedforward prediction. We propose a computational architecture, action space design scheme, and curriculum learning framework that facilitate efficient exploration and sim-to-real transfer.

搜集汇总

背景与挑战

背景概述

KAIST Nonprehensile Objects是一个23.4GB的机器人学习数据集，由韩国科学技术院于2023年发布，专注于基于深度强化学习的非抓握操作系统的研究和应用。