EgoWild2Dex: Training robots with uncurated first-person human data
Overview
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.
From Chinese robotics — Dataset and collection discovery
Event evidence and corrections
1 attributed source owners. Ownership does not establish independent confirmation. Quantities are reported separately and are never added together.
Reported quantity · hours_of_data: 538.9 hours · Basis not reported
“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。”
Exact source · revision 2Source owner not reported
Reported quantity · Reported success rate: 96.7 percent · Basis not reported · Differing source assertions
“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%。”
Exact source · revision 2Source owner not reported
“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%。”
Exact source · revision 2Source owner not reported
“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%。”
Exact source · revision 2Source owner not reported
Reported quantity · head_rotation_per_second: 15.93 other · Basis not reported
“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 34.78 个不同的物体实例。”
Exact source · revision 2Source owner not reported
Reported quantity · object_instances_per_frame: 34.78 other · Basis not reported
“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 34.78 个不同的物体实例。”
Exact source · revision 2Source owner not reported
“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 34.78 个不同的物体实例。”
Exact source · revision 2Source owner not reported
Reported quantity · Reported duration: 1 hours · Basis not reported · Differing source assertions
“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%。”
Exact source · revision 2Source owner not reported
Artifact availability · code: reported pending
“数据集、模型与代码计划后续开放。”
Exact source · revision 2Source owner not reported
“数据集、模型与代码计划后续开放。”
Exact source · revision 2Source owner not reported
Artifact availability · dataset: reported pending
“数据集、模型与代码计划后续开放。”
Exact source · revision 2Source owner not reported
“数据集、模型与代码计划后续开放。”
Exact source · revision 2Source owner not reported
Reported quantity · hours_of_human_data: 538 hours · Basis not reported
“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”
Exact source · revision 2Source owner not reported
Reported quantity · hours_of_real_robot_data: 1 hours · Basis not reported
“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%。”
Exact source · revision 2Source owner not reported
Reported quantity · head_rotation_angle_per_second_degrees: 15.93 other · Basis not reported
“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 34.78 个不同的物体实例。”
Exact source · revision 2Source owner not reported
Environment: controlled real
“EgoWild2Dex 的做法更直接：不再要求人为地把数据「拍干净」，而是直接进入真实环境，记录人类原本就在做的事情。”
Exact source · revision 1jiqizhixin
“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。”
Exact source · revision 1jiqizhixin
“与其先把数据 「清洗」 成实验室里的样子，EgoWild2Dex 选择直接利用保留了视角晃动、复杂背景和长尾物体分布等真实因素的自然数据。”
Exact source · revision 1jiqizhixin
Data origin: human egocentric
“因此，越来越多研究开始尝试利用 人类第一视角（Egocentric）数据 帮助机器人学习。”
Exact source · revision 1jiqizhixin
“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 34.78 个不同的物体实例。”
Exact source · revision 1jiqizhixin
“但这里的「脏」并非指标注错误或数据质量差，而是指它 保留了真实世界本来的复杂性 。”
Exact source · revision 1jiqizhixin
“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？”
Exact source · revision 1jiqizhixin
Reported quantity · Reported duration: 538.9 hours · Basis not reported · Differing source assertions
“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。”
Exact source · revision 1jiqizhixin
Reported quantity · Reported success rate: 96.7 percent · Basis not reported · Differing source assertions
“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%。”
Exact source · revision 1jiqizhixin
Reported quantity · head rotation angle: 15.93 other · Basis not reported
“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 34.78 个不同的物体实例。”
Exact source · revision 1jiqizhixin
Reported quantity · object instances: 34.78 other · Basis not reported
“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 34.78 个不同的物体实例。”
Exact source · revision 1jiqizhixin
Reported quantity · training_hours: 538.9 hours · Basis not reported
“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。”
Exact source · revision 2Source owner not reported
Evidence dependency changes (8)
- 2026-10-30T22:42:14.789Z · evidence updated · source revision 1. Evidence extraction was updated; current source attributions are shown above.
- 2026-10-30T22:42:09.229Z · evidence updated · source revision 1. Evidence extraction was updated; current source attributions are shown above.
- 2026-10-30T22:38:58.059Z · evidence updated · source revision 2. Evidence extraction was updated; current source attributions are shown above.
- 2026-10-30T22:38:42.931Z · evidence updated · source revision 2. Evidence extraction was updated; current source attributions are shown above.
- 2026-10-30T22:33:09.801Z · evidence updated · source revision 2. Evidence extraction was updated; current source attributions are shown above.
- 2026-10-30T22:32:50.619Z · evidence updated · source revision 2. Evidence extraction was updated; current source attributions are shown above.
- 2026-10-30T22:22:05.487Z · evidence updated · source revision 2. Evidence extraction was updated; current source attributions are shown above.
- 2026-10-30T22:21:48.217Z · evidence updated · source revision 2. Evidence extraction was updated; current source attributions are shown above.
Report timeline
Follow attributed reports and material updates.
- Chinese robotics — Dataset and collection discoverySignal538 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
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.
- Chinese robotics — Dataset and collection discoverySignal538 Hours of 'Extremely Dirty' Human Data: Hong Kong University Teaches Robots to Unbox, Use Hot Glue Guns, and Make Cold Drinks
A team from the MMLab at the University of Hong Kong has introduced EgoWild2Dex, a dataset and framework that uses raw, uncurated first-person human operation data to train robots on complex tasks such as unboxing, using a hot glue gun to fix figurines, and making cold drinks. The method achieves an average success rate of 96.7% by employing GeoFormer perspective alignment and action retargeting to transfer natural human data into the robot's action space, reducing reliance on laboratory environments.
- Chinese robotics — Dataset and collection discoverySignal538 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
The MMLab team at the University of Hong Kong proposes the EgoWild2Dex method, which uses uncurated real human first-person operation data to train robots, enabling complex dual-arm dexterous tasks such as unpacking packages, using hot glue guns to fix figurines, and making cold drinks, with an average success rate of 96.7%. The method transfers natural human operation data to robotic execution space through GeoFormer perspective alignment and progressive training.
- Machine Heart — Robotics on WeChatSignal538 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
The MMLab team at the University of Hong Kong proposes EgoWild2Dex, which uses 538.9 hours of uncurated, real-world first-person human operation data to train robots, enabling complex dual-arm dexterous tasks such as unpacking packages, using a hot glue gun to fix figurines, and making cold drinks, with an average success rate of 96.7%. The method aligns human and robot perspectives using GeoFormer and adopts a progressive human-robot joint training paradigm, transforming natural human operation data into executable strategies for robots.
Timezone · UTC
Article dates follow your selected timezone. Briefing editions use Hong Kong time (UTC+8).