Cut VLM round trips in a robot loop with a cheap check first
A small robot asked a remote vision-language model after every 1-2 cm nudge (3-5 s each), which made it stop and go. Two cheap filters removed most of those calls without touching the grasp program itself. (1) Wrap the detect function: on the first call, also find the object's colour blob in the same photo and compare it with the VLM's box (IoU of 0.6 or more). Only if they agree, answer the following calls from the colour blob alone (milliseconds). Fall back to the VLM whenever the blob is missing or the object name changes, and for names with no colour word. (2) Before asking 'is there a white tray in this photo', run a 10 ms mask for white, smooth patches and skip the VLM when there are none. Replaying the saved frames of a failed search confirmed it: 11 of 16 frames were plain carpet and would have been skipped, and every frame showing the tray passed. Check where the time goes from logs first: here the driving was already continuous and the waits were all perception.