EyeRobot 2.0 swaps wrist cameras for active gaze
Justin Kerr asks whether the cameras bolted to almost every bimanual robot are actually necessary.
Justin Kerr (@justkerrding) and @kushtimusprime have released EyeRobot 2.0, a system that throws out the wrist cameras on a bimanual robot and uses active gaze from a single ego stereo camera instead.
The pitch starts with a complaint rather than a capability. Almost all bimanual robots carry wrist cameras, the small cameras mounted near each gripper that give a policy a close-up view of whatever it is about to grab. Kerr's framing is that they are both necessary and annoying.
"They add bulk, cost, and complexity, but are critical for performance. Are they really needed?" he wrote in the announcement.
They add bulk, cost, and complexity, but are critical for performance. Are they really needed?
What active gaze means here
EyeRobot 2.0 replaces that setup with one ego stereo camera, meaning a two-lens head camera that sees from the robot's own point of view and gets depth from the pair. Active gaze means the camera is not a fixed mount. It aims, the way eyes do, pointing at whatever the task currently needs instead of holding one wide view and hoping the resolution is enough.
If that works, the hardware story is simple. No cameras on the wrists means less to cable, less to break on impact, less mass at the end of the arm and one less thing to buy for every new robot. The eye hardware moves to the head, where there is room for it.
What has not been shown yet
Kerr's post is the announcement, and it does not include benchmark numbers, task lists or a comparison against a wrist-camera baseline in the text itself. So treat the headline claim as a design argument so far, not a proven result. The question that matters is whether active gaze matches wrist cameras on the tasks people actually run, or only gets close to them, and that is not answered in the announcement.
It is also worth being clear about what this is not. This is not a new skill or a new task. It is a bet that one well-aimed pair of eyes can do the job of several fixed ones, which is the kind of simplification that only counts if the performance holds. You can read Kerr's thread on EyeRobot 2.0 for the rollout.