FTP-1 Heterogeneous Tactile Manipulation Training Dataset
FTP-1-Dataset 是用于跨传感器触觉 foundation-policy 预训练的公开混合数据分发。官方 Hugging Face 页面可核验约 5.85M 行、7.53TB、18 个 archive 和 WebDataset 格式；论文/项目报告训练混合约 3,000 小时、26 源、21 个触觉传感器，包含人类示范、灵巧手和夹爪机器人。卡片没有 YAML 许可证元数据，未完成全量下载或逐源版权、隐私、动作/标定与成功失败审计，因此它是可访问但使用条件未完成核验的大规模数据入口，而非无条件商用数据集。
- 开放状态
- 数据可下载·许可证缺失
- 许可证
- No dataset license metadata declared on the Hugging Face card; each of 26 upstream sources requires separate license, privacy, consent and redistribution review
- 商业使用
- not established; Apache-2.0 on FTP-1 code does not grant rights to the 7.53TB data mixture, upstream datasets, model checkpoints or robot hardware
- 人工评审结论
- official_public_dataset_viewer_and_file_structure_verified_7_53tb_18_archives_license_missing_upstream_sources_not_fully_audited
- 机器人
- human demonstrations、dexterous-hand systems、gripper-based robots、source-specific tactile manipulation platforms
- 任务
- contact-rich manipulation pretraining across human demonstrations, dexterous hands and gripper robots; full task inventory by archive is not reconciled
- 模态
- tactile observations across image/array/state modalities、vision where present in source collection、robot state、actions、language/instruction fields where provided、source-specific calibration and metadata
- 格式
- Hugging Face WebDataset archive collection; card says each source directory contains a single tar or split tar.part-* files, with concatenation required for split archives
- 采集/生成方式
- aggregation and repackaging of 26 heterogeneous tactile manipulation sources; collection method remains source-specific
- 已知问题
- Card YAML license metadata is missing. Eighteen archive names are not the same as an audited 26-source rights ledger. No full download, file hash ledger, all-shard schema, per-source modality/timestamp calibration, consent/privacy, success/failure balance, action scale or commercial re-distribution check was completed. Do not infer a permissive data license from Apache-2.0 training code or public Viewer access.