Reproduce “INVAR: Randomization-Invariant Visual Representations for Zero-Shot Sim-to-Real” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sunshine Jiang, Shuo Cheng, Francois Robert Hogan, Matthew Chang, Tingfan Wu, Jitendra Malik, Karen Liu, Nur Muhammad Mahi Shafiullah and 1 more
INVAR moves visual domain randomization into reusable encoder pretraining. It aligns clean and randomized renders of the same simulated state using teacher-student self-distillation, then freezes the encoder for policies trained on clean simulation. On bimanual parallel-gripper tasks, mean real-world success was 0.65 versus 0.32 for DINOv3 trained with domain randomization. The result concerns visual robustness; geometry, dynamics, and actuation mismatches remain unaddressed.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.