PolaRiS: Scalable Real-to-Sim Evaluations for Generalist Robot Policies

July 13, 2026

-for sake of the paper, generalist policy = VLA

Steps for constructing the environment:

1.Construct the visuals/geometry of the environment

-2DGS, COLMAP, truncated signed distance field (TSDF)

2.Construct the visuals, geometry, and articulation of the robot

-Use the robot URDF, build Gaussian splats on top of it and assign to each joint

-For every robot movement, rotate the corresponding splats

3.Add object assets to the scene

-Trellis 3D image-to-3D generative model

-PolaRiS allows scene reconstruction/novel view synthesis in 3D, allowing views from any angle (i.e. wrist camera, etc.)

-collect teleop simulation data, mix datasets, and continue behavior cloning on the original VLA pre-training objective