DepthWorld Gives Robots True 3D Vision

🔥 What happened Researchers from Czech Technical University and Inria unveiled DepthWorld, a world model that trains robot manipulation on real 3D geometry instead of RGB-only video. They recalibrated the DROID dataset into DROID-3D, adding dense metric depth and precise multi-view extrinsics. 💡 Why it matters Existing video world models fake consistency: each frame looks fine, but the 3D structure collapses. DepthWorld boosts RGB prediction by +1.48 dB PSNR while delivering accurate metric depth. That's the gap between a pretty video and a usable simulator. ⚡ Our take Finally someone admits it: without 3D geometry, every robot world model is just an expensive screensaver. Those +1.48 dB aren't a marketing stunt — they prove depth sharpens the video prediction itself.