The Humanoid Hub· @TheHumanoidHub · X·· 3 小时前AI 评分48
AI 导读
Perceptron 推出 Egocentric 1.0 自助 API，可将第一人称人类与机器人视频转为结构化训练数据，输出亚秒级子任务时间线、双手（或夹爪）独立活动时间线及无需相机标定的 3D 手部重建。其子任务质量号称媲美 GPT-6 Astra，成本约为其 1/8（每视频小时 9 美元 vs 约 76 美元），全栈方案 29 美元，让人形机器人低成本训练数据的利用更可行。
AI 生成摘要 · 以原文为准
正文
One of the biggest data bottlenecks in robotics is turning raw human video into high-quality training labels.
Perceptron Egocentric 1.0 is a self-service API that turns first-person human and robot video into structured training data:
- Timestamped subtask timelines, down to sub-second actions
- A separate activity timeline for each hand (or gripper)
- State-of-the-art 3D hand reconstruction, no camera calibration needed
It claims to match GPT-6 Astra on subtask quality at ~1/8 the cost ($9 vs ~$76 per video-hour), full stack at $29.
The pipes to leverage the cheapest robot training data just got more viable.
来源：The Humanoid Hub · x.com