Same VLM · Same GPU · Only delta sent when scene is static
Video: 17 sampled frames from 5.4s clip
Fire-and-forget cloud inference — no blocking, results retrieved async. This is how you'd run robot perception at scale in production.