Mehrotra 认为人形机器人内存需求大幅提高,是因为需处理上百个传感器数据并本地运行大模型以实现毫秒级控制。
在他看来,这一差距源于人形机器人需要同时处理上百个传感器(视觉、触觉、力觉等)的数据,并在本地运行大模型以实现毫秒级实时控制,对容量、带宽和延迟提出极高要求。原文 (en)
In his view, the gap stems from the need for humanoid robots to process data from more than a hundred sensors — vision, touch, force and others — simultaneously and to run large models locally for millisecond-level real-time control, placing extremely high demands on capacity, bandwidth and latency.