Robot RL trained on photorealistic Gaussian splats transfers zero-shot

lukas_m_ziegler · x · 2026-07-21

FlexionAI says it has built a robotics RL pipeline that uses real-world site scans and photorealistic Gaussian splats to close the sim-to-real gap.

The team argues that prior robot RL mostly relied on synthetic, untextured environments and depth images because RGB was hard to use in simulation.
With partners Niantic Spatial and NVIDIA Robotics, they scan a real deployment site with off-the-shelf hardware, reconstruct it into a photorealistic Gaussian-splat scene, and train policies at scale in a Gym-like environment.
Those policies then transfer zero-shot to the real robot and the environments they were trained for.
The claim is that this should make deployment faster and policies more capable and robust, with applications beyond navigation.

More from Embodied