Simon Cowling tested four imitation-learning policies on a low-cost SO-101 robot arm doing a pick-and-place task. Each policy was trained on the same 50 demonstrations and tried ten times in a room lit to about 520 lux and ten times again at about 45 lux, which is 11.5 times dimmer.
Success was 20 of 40 attempts in normal light and 19 of 40 in low light, a change of 2.5 percentage points with a wide margin of error. The recorded video explains why: the front camera image was only 3% darker, and the wrist camera image 19% darker.
Both cameras used automatic exposure and white balance, which most likely evened out the difference. The author says exposure was not locked in this experiment. He concludes that light levels in the room do not describe what a policy sees, and that camera settings and image statistics should be reported. Data and code are public.
