538 Hours of 'Extremely Messy' Human Data, Real Robot Takes Less Than an Hour: Hong Kong University Teaches Robots to Unpack Packages, Use Hot Glue Guns, and Make Cold Drinks
538小时「极脏」人类数据，真机不到1小时：港大让机器人学会拆快递、打胶枪、做冷饮
The MMLab team at the University of Hong Kong proposed the EgoWild2Dex method, using uncurated real human first-person operation data to train robots, successfully completing complex tasks such as unpacking packages, using a hot melt glue gun to fix figurines, and making cold drinks, with an average success rate of 96.7%. The study aligns human and robot perspectives through the GeoFormer module and adopts a progressive human-robot joint training paradigm, unifying open scene data, task-related data, and real robot data into the robot's action space.
The study demonstrates that raw, uncurated human-first-person data can be effectively used to train robots for complex tasks, with a novel approach to align human and robot perspectives through geometric transformations and progressive training.
Source: Chinese robotics — Dataset and collection discovery · Read original article ↗
English translation unavailable. Choose the original article to read it.
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
training_hours
538.9 hours
Reported success rate
96.7%
View original evidence
在每个下游任务使用不到 1 小时真机数据的情况下，机器人能够连续完成拆快递、使用热熔胶枪固定手办、制作冷饮等复杂双臂灵巧长时序任务，平均成功率达到 96.7%。
Open source S5
- code: not_reported
- dataset: not_reported
- weights: not_reported
- paper: not_reported
What remains unknown
Not established in the collected evidence: Environment, Control, Data origin.
Reported performance applies to the described task. It does not establish general autonomy or deployment readiness.
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
近日，团队提出 EgoWild2Dex，直接利用 538.9 小时未经刻意编排的真实人类第一视角操作数据训练机器人。
Open source S4
（来源：机器之心Pro） 视频链接：https://mp.weixin.qq.com/s/01a9kUTeDlGKkUH07C2CWQ 如果给机器人看的训练视频，来自杂乱的家庭、繁忙的工厂和药房，相机还随着人的走动不断剧烈晃动，它还能从中学会精细操作吗？
Open source S2
在每个下游任务使用不到 1 小时真机数据的情况下，机器人能够连续完成拆快递、使用热熔胶枪固定手办、制作冷饮等复杂双臂灵巧长时序任务，平均成功率达到 96.7%。
Open source S5
数据显示，EgoWild 中，每秒累计头部旋转角度平均达到15.93°，每帧图像平均包含个不同的物体实例。
Open source S16
按传统机器人数据采集的标准来看，这批数据确实相当「脏」。
Open source S17
Source：Chinese robotics — Dataset and collection discovery · sohu.com