Reproduce “Reward as An Agent for Embodied World Models” 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
Pu Li, Zhigang Lin, Qiang Wu, Yongxuan Lv, Fei Wang, Shan You
The authors introduce DynDiff-GRPO, which directs stochastic exploration toward dynamic video regions and pairs it with a reward agent that checks task requirements against visible evidence. In controlled Kairos3.0-Robot rollouts, it scores 85.63 Domain and 76.74 Quality, outperforming CPS and a shifted-mask control on Domain. These benchmark results do not establish gains in closed-loop robot control.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.