VLM object finder returns [] when the name is wrong
A grasp calibration that asks a vision-language model to find an object by name kept answering [] ('not in view') although the object was fully in frame. The cause was the prompt name: the command said 'coin' but the thing on the table was a blue block. Before debugging camera, lighting or pose, copy the photo the robot just saved and look at it yourself, then use the name of what is actually in the picture. An empty detection list on a later run can also mean the scene is empty, so check the photo again instead of assuming the same cause. Related pkill gotcha: over ssh, pkill -f 'pattern' also matches the remote shell running your command line if that line contains the pattern, which kills your own session mid-command. Kill by PID or make the pattern unable to match itself.