Overcoming RGB Sim-to-Real Gap: Gaussian Splatting Enables Zero-Shot Robot Transfer

ZeYanjie · x · 2026-07-23

Traditional robot RL policies are often trained in synthetic, untextured environments, forcing perceptual policies to rely on depth images due to a manageable sim-to-real gap. Leveraging RGB in simulation has remained an open problem.

Partnering with Niantic and NVIDIA, FlexionAI built a pipeline to address this. They scan real deployment sites with off-the-shelf hardware, reconstruct them into photorealistic Gaussian splats, and run massively parallel RL training. The resulting policies transfer zero-shot to real robots, enabling faster deployment.

Original post →

More from Embodied

Embodied channel →