Cosmos3-DROID
收藏资源简介:
DROID(分布式机器人交互数据集)是一个大规模“野外”机器人操作数据集,旨在通过大幅提升场景、任务、物体、视角和交互地点的多样性,训练可泛化的机器人操作策略。数据集包含约7.6万条遥操作演示轨迹(约350小时交互数据),在12个月内采集于13个机构的18个实验室,覆盖564个独特场景、86种任务和52栋建筑。数据使用统一的开源机器人硬件平台(Franka Panda 7-DoF机械臂与Robotiq 2F-85夹爪)收集,确保跨实验室数据的可比性和可复现性。每个数据片段包含三个同步的立体RGB相机视频流、相机标定数据、深度信息、底层机器人状态与控制命令,以及最多三条自然语言任务指令。数据以LeRobotDataset v3.0标准组织,包含成功和失败两个独立子集,总计约7.2万条片段,超过2240万帧数据,总存储量约707 GB。特征包括来自三个视角(两个外部、一个腕部)的视频观察(640x360 RGB,15 FPS)、7维关节位置/速度/扭矩状态、6维笛卡尔位置状态、夹爪状态,以及相应的关节/笛卡尔/夹爪位置与速度动作。数据集适用于机器人策略学习、机器人前向动力学模型学习和逆向动力学模型学习等任务,研究表明其能提升策略性能、鲁棒性和泛化能力。基于OpenMDW1.1许可证发布,可用于商业或非商业用途。
DROID (Distributed Robot Interaction Dataset) is a large-scale "in-the-wild" robotic manipulation dataset developed to train generalizable robotic manipulation policies by substantially enhancing the diversity of scenarios, tasks, objects, viewpoints, and interaction locations. It contains approximately 76,000 teleoperation demonstration trajectories (roughly 350 hours of interaction data), collected over a 12-month period across 18 laboratories affiliated with 13 institutions, covering 564 unique scenarios, 86 distinct tasks, and 52 buildings. All data was collected using a unified open-source robotic hardware platform (Franka Panda 7-DoF robotic arm paired with a Robotiq 2F-85 gripper) to guarantee cross-laboratory data comparability and reproducibility. Each data segment includes three synchronized stereo RGB camera video streams, camera calibration parameters, depth information, low-level robot states and control commands, and up to three natural language task instructions. The dataset is structured per the LeRobotDataset v3.0 standard, and comprises two independent subsets: successful and failed trials, with a total of approximately 72,000 segments, over 22.4 million frames of data, and an overall storage footprint of around 707 GB. The collected features include video observations (640×360 RGB, 15 FPS) from three viewpoints (two external, one wrist-mounted), 7-dimensional joint position/velocity/torque states, 6-dimensional Cartesian position states, gripper states, as well as corresponding joint, Cartesian, and gripper position and velocity actions. This dataset is suitable for tasks including robotic policy learning, robotic forward dynamics model learning, and inverse dynamics model learning. Existing research has demonstrated that it can improve policy performance, robustness, and generalization ability. The dataset is released under the OpenMDW 1.1 license, and permits both commercial and non-commercial usage.




