yeeeiii111/wuji-pick-and-place
收藏资源简介：
Wuji拾放捆绑数据集是一个用于双手拾放任务的配对人类演示与机器人遥操作数据集。该数据集旨在训练人类到机器人转移策略，例如基于第一人称人类参考视频的条件动作扩散模型。数据集包含三个子集：teleop（机器人遥操作数据，179个片段，使用三台相机：头部和左右手腕相机，分辨率分别为720×1280和480×848）、ego_ref（人类自我中心参考视频，179个片段，与teleop一一配对，使用头部相机，分辨率720×1280）和eval_refs（评估用参考视频，102个片段，包含域内剩余数据和域外数据）。数据集采用LeRobot v2.1格式，包含8个域内任务（任务索引0-7），涉及不同物体（球、立方体、杯子）和抓取方式（抓取身体或把手）。数据集还提供了状态和动作数据（54维），以及任务提示模板。
The Wuji Pick-and-Place Bundle is a paired human demonstration ↔ robot teleoperation dataset for bimanual pick-and-place tasks. It is built to train human-to-robot transfer policies, such as VAM-style action diffusion conditioned on first-person human reference video. The dataset consists of three subsets: teleop (robot teleoperation data, 179 episodes, with three cameras: head and left/right wrist cameras at resolutions 720×1280 and 480×848), ego_ref (human egocentric reference videos, 179 episodes, one-to-one paired with teleop, using a head camera at 720×1280 resolution), and eval_refs (evaluation-only reference videos, 102 episodes, including in-domain leftover and out-of-domain data). The dataset is in LeRobot v2.1 format and includes 8 in-domain tasks (task_index 0-7) involving different objects (ball, cube, cup) and grasping methods (by the body or by the handle). It provides state and action data (54-dimensional) and task prompt templates.
