Ep#62: PolaRiS: Scalable Real-to-Sim Evaluations for Generalist Robot Policies

Evaluating robotic policies in the real world is notoriously difficult, slow, and non-reproducible, yet simulation benchmarks often fail to capture real-world performance accurately. A new approach addresses this by leveraging 3D Gaussian splatting and generative models to create scalable, low-cost simulation environments that serve as reliable proxies for real-world tasks. By incorporating a small amount of "sim co-training" data, these models learn to ignore spurious visual artifacts, significantly improving the correlation between simulated and real-world outcomes. Unlike fully data-driven video models that struggle with long-horizon consistency and failure representation, this hybrid method maintains classic physics while providing high-quality visuals. This framework enables researchers to rapidly iterate on policies and crowdsource diverse evaluation environments, ultimately aiming to establish a standardized, community-driven benchmark that pushes robotics research toward more robust, generalizable performance.