Real Robot Data Strategy

Real robot data strategy is the approach to robot model training described by Gao Jiyang in 132. 对星海图创始人高继扬的3小时访谈：鲶鱼、曾国藩、Waymo与Momenta的两面、一只狼与许华哲的离开. He argues that useful embodied intelligence should be trained as much as possible on target-domain data, while still experimenting with simulation, teleoperation, human-centric data, point-of-view data, third-person video, and other sources.

134. 【数据的综述】和谢晨聊，新时代的石油、历史、版图、数据金字塔、定价与Recipe qualifies this view through 谢晨. He agrees that real robot data is valuable, but argues it can be overestimated when treated as the main scalable path; in his Embodied Data Pyramid, real robot data is the top layer because it is accurate, expensive, and hard to scale.

170: 【具身季报 26Q2】世界模型大风不停，和不想被贴标签的人 adds Embodied Robot Data Paradigms as the time-varying version of the data problem. Chen Zhe Peter traces shifts from Aloha-style real-robot teleoperation to UMI body-free collection, egocentric video, whole-body motion capture, and dexterous-hand data, arguing that each model-route change depends on a new way to collect relevant physical experience.

从会跳舞到有感知，触觉是机器人通往智能的门票吗？｜ S10E19 adds Eric Li Zhiqiang / 李志强’s tactile-data version. He says real machine data is best but may only be 10-20% of the recipe because collection is costly and scarce; Yimu Technology / 一目科技 therefore combines real Tactile Sensing data, simulation with Optical Tactile Sensing, and large-scale video pretraining before aligning touch with robot actions.

166: 许华哲再次具身创业：不想错过最大的西瓜 adds Xu Huazhe’s household-robot version. He expects more video data to enter robot training, says teleoperation can show progress but may not be the final data source, and argues that failure data or suboptimal data should be used selectively rather than discarded or mixed blindly.

E244｜端到端vs上下分层：机器人路径之争，正在转向？ adds Han Zheng / 韩正’s sharper critique of real-data scaling. He argues that robots do not yet have a Tesla-like deployed fleet, and that asking users to teleoperate household robots at massive scale is not a realistic substitute. The source therefore treats real robot data as necessary validation and adaptation data, not as the sole source of Open-World Robot Manipulation.

146. 对Physical Intelligence柯丽一鸣4小时访谈：Pi的开源模型研究，机器人的江湖、族谱与主角 adds K’s Robot Experience Data distinction. Human teleoperation data can start a policy, but π0.6* uses robot-owned attempts, failures, and correction traces so Robot Reinforcement Learning can improve task throughput on real machines.

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 沈宇军’s 蚂蚁灵波 version. Shen favors real-machine data for training because it contains real sensor noise, embodiment, and execution distribution, but he also says scale depends on task design, cross-body cleaning, first-person human data, and whether the resulting data is usable by Embodied Native Foundation Models.

Key Claims

• Traditional graphics simulation can have a large sim-to-real gap, so it should not be assumed to replace real robot operation data.
• Data cost has to be counted together with training cost and engineer cost; low-quality data can waste the expensive parts of the stack.
• The right “data recipe” is empirical: different data types may help, but their proportions have to be discovered through experiments.
• Scaling real data requires entering real scenes and distributing collection devices or robots widely enough for the data to compound.
• Robotics lacks the autonomous-driving-style installed fleet that would make passive real-world data collection cheap and broad.
• Dexterous-hand data is especially body-specific: hand geometry, motors, degrees of freedom, and sensors can make retargeting across hardware difficult.
• Tactile data adds force, deformation, friction, and slip signals that visual data does not contain, but it must be processed quickly enough for real-time correction.
• Unified Robot Models require data selection, not only data volume, because post-training can otherwise improve fixed tasks while shrinking generalization.
• Robot Experience Data is valuable because it binds action, failure, correction, and embodiment in the robot’s own hardware rather than only in human demonstrations.
• Robot Data Scale Up is not raw recording volume: camera placement, body configuration, task coverage, data cleaning, and model usability decide whether additional hours improve generalization.

Connections

• Physical World Data Flywheel — larger loop that turns data into product improvement.
• Xinghaitu, Gao Jiyang, and Vision Language Action Models — source company, speaker, and model route.
• Causal World Models and World Models — adjacent model directions where data quality and physical grounding also matter.
• Embodied AI — broader field where real-world distribution shift makes data strategy central.
• 谢晨, 光轮智能, Embodied Data Pyramid, and Robotics Simulation Evaluation — source and concepts that put real robot data inside a simulation-centered data loop.
• Embodied Robot Data Paradigms, Robot Teleoperation and Remote Takeover, Dexterous Manipulation, and Generalist — new data-collection and body-specific-data layer from the LateTalk source.
• Yimu Technology / 一目科技, Tactile Sensing, Optical Tactile Sensing, TouchNet, and Tactile Transformer Encoder — tactile real-data and model-interface layer added by the What’s Next source.
• Poke Robotics, Xu Huazhe, AI Native Robotics, Unified Robot Models, and Robot Active Use Metrics — household-robot data route added by episode 166.
• 速度科技 / Sudu Technology, Structured 3D Robot Data, Sim2Real, and Open-World Robot Manipulation — E244’s critique of real-data-only scaling.
• Physical Intelligence, Physical Intelligence Pi Model Series, Robot Experience Data, and Robot Reinforcement Learning — experience-data and performance-improvement route added by episode 146.
• 蚂蚁灵波 / Ant Lingbo, 沈宇军 / Shen Yujun, Robot Data Scale Up, and Embodied Native Foundation Models — real-machine-heavy data-scaling route added by episode 147.