Camera-only table-edge detection: learn the surface's hue once, sample the surface per frame (auto white balance breaks fixed colour models)
Follow-up to a camera-only 'stay on the table' check for a small robot (no cliff sensors). Version 1 learned the table's Lab colour clusters once at the start and compared every later frame to them. On a dark glossy wooden table it failed badly: the robot's USB camera has auto white balance / exposure, so the SAME table top came out purple in one view and orange in another, plus big specular glare spots; only 1 of 12 real photos was judged 'table', so the robot would refuse to drive almost everywhere.
Version 2 (judged 10 of 12 on the same photos): - Per frame, take the strip just ahead of the robot as the surface sample of THIS frame (same white balance and light as the rest of the frame). - Accept that sample only if (a) its tone matches the learned table tone in hue angle (+-25 deg) and chroma (0.4-2x) - brightness is NOT compared - and (b) it is one uniform surface: robust colour spread (MAD of a, b in Lab) <= 6 and brightness spread <= 30. Real table strips had colour spread <= 4.4; another robot's arm in front of the lens 22-30; a laptop 9. Neutral tile floors, door mats, paper fail the tone test (for a neutral-coloured table, fall back to chroma/brightness). - Table = pixels like the sample (colour tight, brightness loose x2.5) connected to it, plus everything the region encloses (glare spots, objects on the table) via a flood fill of the outside from the image border. - Implausibility guard: if 'table' covers the top rows of a downward-looking view (~2 m away or above the horizon), something is right in front of the lens: treat as not judged (no forward motion).
Also: the camera's nearest visible ground was ~15 cm past the bumper, so an object placed right in front of the robot was invisible to the search; on a table you cannot back up blindly to see it. Use a pose that looks straight down at the bumper (no driving) to check that blind spot before declaring 'nothing left'.