Third-Person Camera

A camera mounted off the robot’s body, fixed in place, viewing the whole workspace from the side.

A third-person camera is mounted on a tripod, wall, or table edge and doesn’t move with the robot arm, in contrast to a wrist camera or head camera. It has a wide field of view and captures the robot, objects, and environment together, giving global position information; the tradeoff is that distant details are hard to see and the arm can easily occlude objects, so manipulation datasets usually pair it with a wrist camera. This setup is a form of “eye-to-hand” mounting, and its extrinsics — the camera’s position and orientation relative to the robot — must be calibrated before image coordinates can be converted into the robot’s base frame. If the camera moves even slightly, the scene a policy sees changes, and success rates often drop, which is why generalization across viewpoints is a common evaluation axis. In papers this camera is also called an exterior camera or 3rd-person camera.

ExampleEach capture rig in the DROID dataset has two repositionable ZED 2 stereo cameras serving as third-person views, plus a ZED Mini mounted on the wrist; every time the scene changes, the operator repositions the third-person cameras and recalibrates their extrinsics with a checkerboard.

Also called

Static Camera, External Camera, Exterior Camera, Global Camera, 3rd-Person Camera

Related

Wrist Camera · Head Camera · Eye-to-Hand · Camera Extrinsics · Multi-View · DROID (Distributed Robot Interaction Dataset)

Sources

DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset