ManiSkill
收藏资源简介:
ManiSkill是一个大规模的开源基准,用于从3D视觉输入学习物理操纵技能。该数据集由加州大学圣地亚哥分校创建,包含约36,000条成功轨迹,总计约150万点云/RGB-D帧。数据集通过模拟全景相机捕捉的自我中心点云或RGB-D图像,支持多种计算机视觉模型应用。创建过程中,通过精心设计的共享奖励模板自动生成奖励函数,使用强化学习方法收集大量成功轨迹。ManiSkill旨在促进学习-从-示范方法的发展,并解决家庭环境中的多样化操纵挑战。
ManiSkill is a large-scale open-source benchmark for learning physical manipulation skills from 3D visual inputs. Developed by the University of California, San Diego, the dataset contains approximately 36,000 successful trajectories, totaling roughly 1.5 million point cloud/RGB-D frames. It provides egocentric point clouds or RGB-D images captured by simulated panoramic cameras, supporting applications across a wide range of computer vision models. During its creation, reward functions were automatically generated via meticulously designed shared reward templates, and a large corpus of successful trajectories was collected using reinforcement learning methods. ManiSkill aims to facilitate the development of learning-from-demonstration approaches and address diverse manipulation challenges in household environments.

- 1ManiSkill: Generalizable Manipulation Skill Benchmark with Large-Scale Demonstrations加州大学圣地亚哥分校 · 2021年



