DeVA YAM Real Bimanual Dataset

DeVA作者公开的YAM真实双臂处理数据：90个episode、30,749帧、三路同步RGB、14维动作与本体状态，并附affordance和depth伪标签。固定Hugging Face revision公开非gated、20.44GB；规模小且失败轨迹、隐私与原始标定口径未披露。

开放状态: 数据可下载·上游条款待逐项确认
许可证: MIT declared by dataset card; YAM platform, UAD and DepthAnything terms remain separate
商业使用: dataset card permits MIT use, subject to upstream platform and pseudo-label dependencies
人工评审结论: verified_public_ungated_real_bimanual_processed_dataset
机器人: I2RT YAM 14-DoF bimanual robot
任务: handover marker、lift pot、pick up diverse bottles
模态: three tiled RGB views、14-D action、14-D proprioceptive state、language instruction、affordance pseudo-label、relative-depth pseudo-label
格式: MP4, JSON, NPZ and precomputed T5-XXL embeddings
采集/生成方式: real YAM bimanual execution; collection interface and teleoperation/autonomy split not disclosed on dataset card
已知问题: Only 90 episodes; no failure balance, privacy treatment, raw calibration, checksum audit or independent downstream reproduction; affordance/depth are pseudo-labels rather than physical sensors.