Embodied Robot Data Paradigms
Embodied robot data paradigms are the changing collection methods behind robot model progress in 170: 【具身季报 26Q2】世界模型大风不停，和不想被贴标签的人. Chen Zhe Peter says each model-paradigm shift tends to follow a data-paradigm shift, and the episode traces a path from Aloha-style real-robot teleoperation to UMI body-free collection, first-person video, whole-body motion capture, and dexterous-hand data.
147. 和蚂蚁灵波沈宇军聊：机器人原生基础模型、大脑和本体的关系、预训练与数据scale up、老师汤晓鸥 adds 蚂蚁灵波’s data-native turn. 沈宇军 says the hard part is not only choosing a collection device, but deciding what tasks to assign, how to cover head/waist/base/hand degrees of freedom, and how to clean data across bodies with different cameras and kinematics.
This concept extends Real Robot Data Strategy and Embodied Data Pyramid. It does not say one data type replaces all others; instead, it asks which data source makes a specific capability newly learnable and transferable to a given robot body.
Key Claims
- Whole-body motion capture makes locomotion and manipulation data easier to scale when hardware such as Unitree Robotics becomes a common research platform.
- Dexterous-hand data is highly hardware-specific because finger layout, degrees of freedom, motors, and sensors affect retargeting.
- UMI-style and egocentric data can broaden scene and task coverage, but robot-body deployment remains necessary for final grounding.
- Cross-embodiment cleaning becomes its own paradigm problem when the same model is expected to learn across mobile bases, wrists, heads, waists, grippers, and dexterous hands.
Connections
- Real Robot Data Strategy, Embodied Data Pyramid, and Physical World Data Flywheel — adjacent data-loop concepts.
- Dexterous Manipulation — hardware-specific data case.
- Generalist and Genesis Robotics — companies discussed through large interaction or dexterous-operation data claims.
- 蚂蚁灵波 / Ant Lingbo, 沈宇军 / Shen Yujun, Robot Data Scale Up, and Embodied Native Foundation Models — data-native and cross-body cleaning update from episode 147.