独立行业技术评审 · 持续核验

RoboNet Real Multi-Robot Manipulation Dataset

RoboNet 是早期跨机构真实机械臂数据基础设施:约 162,000 条视频/动作轨迹、超过 1,500 万帧、7 种机器人、113 个相机视角和 4 家机构。它支持视觉前瞻与监督逆模型,并在持出 Franka/Kuka 上评估跨本体预训练与微调。官方项目页声明公开数据,仓库将数据 CC-BY-4.0 与代码 MIT 分离;它是 BridgeData、Open X 与跨本体 VLA 数据谱系的前序节点,不是人形、足式、触觉或全身控制数据集。

RoboNet Real Multi-Robot Manipulation Dataset

RoboNet 是早期跨机构真实机械臂数据基础设施:约 162,000 条视频/动作轨迹、超过 1,500 万帧、7 种机器人、113 个相机视角和 4 家机构。它支持视觉前瞻与监督逆模型,并在持出 Franka/Kuka 上评估跨本体预训练与微调。官方项目页声明公开数据,仓库将数据 CC-BY-4.0 与代码 MIT 分离;它是 BridgeData、Open X 与跨本体 VLA 数据谱系的前序节点,不是人形、足式、触觉或全身控制数据集。

开放状态
已核验开放
许可证
CC-BY-4.0 for the RoboNet dataset as declared by the official repository; repository code is separately MIT
商业使用
dataset attribution obligations apply; separately review code, robot drivers, upstream assets and any model or derivative-data terms
人工评审结论
official_project_paper_and_repository_verified_public_data_with_separate_code_data_licenses
机器人
7 heterogeneous robot-arm platforms including Franka and Kuka
任务
vision-based robot-object interaction, visual foresight, supervised inverse models and cross-robot transfer
模态
robot-object interaction video、action sequence、camera-view metadata、robot/platform identifiers、end-effector state/action representation
格式
official loader/training code and TensorFlow Datasets documentation describe the dataset; exact native archive layout and all split manifests were not downloaded in this review
采集/生成方式
largely autonomous self-supervised real-robot collection with minimal human intervention, pooled across multiple institutions and robot platforms
已知问题
Historical arm-only corpus; no humanoid, legged, whole-body, tactile or force channel is established here. Full archive/checksum, action schema for every native robot, calibration/timing ledger, retained failure proportion, privacy documentation, current dependency compatibility and independent physical reproduction remain unverified.

PMLR 原文 · 官方项目与数据页 · 官方代码与数据许可边界 · TensorFlow Datasets 表示说明