MuJoCo robosuite simulation environments

该数据集包含了针对各种机器人操作任务模拟环境的集合，如螺母插入、两件组装、咖啡制作、螺母和钉子组装以及物理块抓取等。这些任务都具有丰富的接触操作，且在对象几何形状、姿态、观察误差以及操作阶段上各有不同。此外，该数据集不仅包括人类演示，还包含了使用I-Gen系统生成的合成干预，旨在提高策略在面对分布偏移时的鲁棒性。规模上，它将10次人类干预扩展到了1000次合成干预。这些任务旨在模拟和物理环境中执行机器人操作任务。This dataset comprises a collection of simulated environments for various robotic manipulation tasks, including nut insertion, two-part assembly, coffee preparation, nut and nail assembly, and physical block grasping. These tasks feature contact-rich manipulation and vary in terms of object geometry, pose, observation errors, and operational stages. Furthermore, this dataset not only includes human demonstrations but also contains synthetic interventions generated using the I-Gen system, which aims to enhance the robustness of policies against distribution shifts. In terms of scale, it expands from 10 human interventions to 1000 synthetic interventions. These tasks are designed to simulate robotic manipulation tasks that can be executed in physical environments.

搜集汇总

背景与挑战

背景概述

该数据集基于MuJoCo robosuite仿真环境，用于支持IntervenGen系统，该系统通过自动生成干预数据来增强机器人模仿学习的鲁棒性和数据效率。数据集包含模拟和真实机器人实验，涵盖多种任务如立方体抓取和组件装配，并展示了在姿态噪声和分布偏移下的策略恢复能力。