Arhan, a PhD student at the University of Washington, and Carl, a robotics researcher at Physical Intelligence, discuss PolaRiS and how it uses real-to-simulation methods to evaluate robot policies with RoboPapers’ Chris and other guests.
Arhan says existing simulation benchmarks can be difficult to build and may not support moving wrist cameras. Carl explains that real robot tests are slow and hard to reproduce, while simulations are easy to repeat but often fail to predict real-world performance. PolaRiS aims to make simulation scores track how policies perform on real robots.
The team reconstructs environments with 2D Gaussian Splatting for meshes and Gaussian Splatting for visuals, then uses generative models such as TRELLIS to create objects whose hidden sides cannot be scanned. Policies transferred directly from real data showed weak real-to-sim correlation. Adding a small amount of simulation data from unrelated tasks improved alignment, helping policies handle visual differences such as fuzzy close-up views.
The experiments focused mainly on rigid-object tasks and used estimated default object properties rather than extensive physics identification. A hardware-in-the-loop test helped separate visual effects from simulated dynamics; the discussion points to visual mismatch as the larger source of correlation problems for these tasks. Video-model rollouts could produce plausible behavior, but object disappearance and a bias toward successful outcomes made their scores less predictive.
The guests say benchmarks need both community adoption and enough difficulty to keep distinguishing policies as they improve. PolaRiS can help test generalization to unfamiliar objects and scenes, and its browser tool lets users assemble scanned or existing assets before defining success conditions. The team hopes contributors will share environments through the PolaRiS Hub on Hugging Face, then use simulation for iteration before real-world evaluation.
The current approach is aimed mainly at pick-and-place and some articulated-object tasks; deformable manipulation remains difficult to simulate and validate. The speakers see video models and richer physics simulation as two possible paths for handling tasks such as cloth folding. Supporting additional robots also depends on having robot models and sufficiently diverse open datasets for training and comparison.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.
Jang Ren, a Peking University pre-PhD researcher collaborating with Galbot, Tsinghua University, SG2, Calia, and NVIDIA, discusses LDA-1B and its approach to training robot action models from mixed embodied data with RoboPapers hosts Chris and Michael.
Jason Ma, a robotics researcher at the University of Pennsylvania who worked on the paper at Google DeepMind, joins RoboPapers hosts Michael and Chris to discuss using vision-language models to assess robot task progress and data quality.
