Duke just made a pretty compelling case for why humanoid robots shouldn't blindly copy human anatomy. Meet Duke Humanoid V2, a fully open-source 31-DoF humanoid with two RGB-D cameras that can move and look independently. The idea is simple: A robot being able to physically reach something doesn't mean it can actually see it while manipulating it. So the team introduced a new metric called Visible-Reachable Workspace, measuring where the robot can both reach AND perceive. The results are impressive! On the same robot, actuating the cameras increased visible-reachable workspace from 38% to 97%. Adding a second independently controlled camera increased concurrent two-region coverage from 0.45 to 0.95. A third camera only pushed it to 0.97. And in two-target manipulation tests, the dual-camera setup completed tasks 17% faster while consuming 19% less mechanical energy than the same robot with fixed cameras. Bad perception geometry doesn't just create a perception problem. The robot has to compensate physically by turning, repositioning or walking to acquire another view. Sensor architecture therefore becomes part of the robot's mechanical efficiency. At RoboIndex, we track payload, reach, battery life, hand DoF, speed and more when comparing humanoids. The work by @Boyuan__Chen and the Duke General Robotics Lab made me think: maybe we should start tracking how much of their workspace they can actually see too. 🎥 Credits: @Boyuan__Chen