RGB Camera

An ordinary camera that outputs color images, recording red, green, and blue brightness at every pixel.

An RGB camera is simply the most common kind of color camera: its image sensor is covered by a color filter array, most commonly the Bayer filter patented by Kodak engineer Bryce Bayer in 1976, in which half the pixels sense green and a quarter each sense red and blue; a demosaicing algorithm then interpolates a full R, G, B value for every pixel. RGB cameras are cheap, high-resolution, and information-rich, making them the main input for most imitation-learning policies and VLA models. Their limitation is that a single RGB (monocular) camera only captures a 2D projection of the 3D world and has no direct sense of how far away things are; getting depth requires a depth camera, a stereo camera, or a monocular depth estimation model. Other factors to weigh when mounting one include field of view, frame rate, and whether it uses a global shutter (the whole frame exposed at once) or a rolling shutter (exposed row by row, which can skew fast-moving subjects).

ExampleThe original ALOHA dual-arm platform is fitted with 4 ordinary webcams: two mounted on the wrists, one facing forward, and one overhead looking down. ACT, a policy that uses only these color images plus joint angles — no depth — learned to open a semi-transparent condiment cup and insert a battery from about 10 minutes of demonstrations, reaching 80–90% success.

Also called

Color Camera, Monocular Camera

Related

Depth Camera · Stereo Camera · Wrist Camera · Monocular Depth Estimation · Camera Intrinsics · Global Shutter / Rolling Shutter

Sources

Wikipedia: Bayer filter
ALOHA / ACT project page (Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware)
arXiv 2304.13705: Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware