Reproduce “PrimeSim: A Generalist World Model Simulator from Large-Scale Cross-Embodiment Demonstrations” 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
Travis Davies, Wentao Ma, Lihao Guo, Tengfei Ma, Xiaowen Chang, Jiafeng Xu, Shilong Wang, Yiqi Deng and 5 more
PrimeSim turns robot actions into synchronized video views and extends them through self-conditioned rollouts. It trains on generated histories and uses a bounded cache for initial scenes, sparse older context, and recent motion. On RoboTwin 2.0 evaluations, hierarchical memory lowers PSNR degradation from 8.427 dB to 6.008 dB versus a rolling cache. Policy evidence is from simulation, not physical-robot improvement.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.