BridgeData V2 Real-World WidowX Manipulation Dataset
BridgeData V2 是面向可扩展机器人学习的真实跨机构操作数据：官方页面披露 60,096 条轨迹、24 个环境和 13 类技能，其中 50,365 条为 VR 遥操作、9,731 条为脚本化 pick-and-place rollout。每条轨迹配自然语言任务标注；数据来自低成本 WidowX 250 6-DoF 机械臂，使用固定 RGB-D、两路随机 RGB 和腕部 RGB。数据本体采用 CC-BY-4.0，并有官方下载入口；它对 VLA、跨本体操作和离线学习有高迁移价值，但不是人形、足式或全身控制数据。
- 开放状态
- 已核验开放
- 许可证
- CC-BY-4.0 declared by the official BridgeData V2 project page for the data; code, pretrained weights, robot drivers and third-party dependencies have separate terms
- 商业使用
- CC-BY-4.0 attribution obligations apply to the data; verify software, model, robot, sensor and any derivative-data terms separately before commercial deployment
- 人工评审结论
- official_project_and_paper_verified_public_data_download_cc_by_4_0
- 机器人
- WidowX 250 6-DoF robot arm
- 任务
- goal-image and language-conditioned object manipulation, pick-and-place, pushing, sweeping, door/drawer interaction, stacking, cloth folding and granular-media sweeping
- 模态
- natural-language task labels、goal/initial/final visual states、fixed over-the-shoulder RGB-D、two randomized-pose RGB cameras、wrist-mounted RGB、robot manipulation trajectory
- 格式
- official download is described as JPEG archives; exact action, metadata, split and calibration file schemas need archive-level inspection before conversion claims
- 采集/生成方式
- VR-controller teleoperation for demonstrations plus a scripted pick-and-place policy for the autonomous rollout subset; camera positions were expanded during collection
- 已知问题
- Not humanoid, legged, whole-body, tactile or force-labelled. Most data only has the primary fixed camera because extra cameras were added gradually; not every episode has all four views. No full download/hash was verified here; public data license must not be extended to code, weights, drivers or dependencies; no independent physical reproduction was established.