MaxGlove 采集手部力数据，补上机器人训练缺失的一环
一家叫 Maxinsights 的公司做了只手套，记录人手抓东西时的力度数据，专门补摄像头拍不到的那部分，给机器人训练用。
Maxinsights 推出 MaxGlove，一种在人手工作时记录触觉和压力数据的设备。相机只能捕捉抓握的形状，而力信号决定了物体是滑落还是被捏碎，滑落时的修正往往从指尖开始，早于画面上可见的变化。仅用视频训练的策略必须推断力的大小，手套数据则直接提供这一信号，且来自真实家用场景中人们操作物体的过程。Maxinsights 已积累超过 200 万小时视频，目标是 2027 年前达到 100 万小时力数据。
Homes are becoming one of the richest places to collect robot training data, but its not as simple as just recording a video.
Hand force data is becoming one of the most useful signals in robot manipulation.
MaxGlove, from @Maxinsights records touch and force from people’s hands while they work, the layer of a grasp a camera can’t capture.
Pressure decides most of a grasp, whether the glass slides out or cracks. The correction when something starts to slip begins in the fingertips, usually before a camera can pick it up.
A policy trained only on video learns the shape of a grasp and has to infer the force. Glove data supplies the force directly, from people handling the objects robots will be asked to handle.
Maxinsights says it’s aiming for a million hours of force data by 2027, on top of the 2M+ hours of footage captured so far.
