StereoEngine: Scaling Stereo Bimanual Robot Data without Stereo Cameras

This research introduces StereoEngine, a method to synthesize stereo data for robotic manipulation without real stereo cameras. It proposes two complementary approaches: one that warps monocular video into stereo pairs using depth estimation and video diffusion models, and another that generates photorealistic stereo views in simulation. The results show that StereoEngine improves the quality of synthetic data and enhances downstream bimanual manipulation tasks.