Perceptron says its new early-access product, Egocentric, turns robot-camera and first-person video into policy-training-ready supervision by segmenting manipulation events, generating labels and tracking each hand through an episode. In posts on X, the company and supporters said the system delivers per-frame hand boxes, 21-keypoint skeletons, persistent left-right identity and per-hand action captions, and that it reached a 0.280 semantic end-to-end F1 on WGO-Bench versus 0.158 for prior Gemini 3.5 Flash and Gemini Robotics-ER 1.6 based annotation pipelines, according to Scobleizer's summary of the launch thread.