Reproduce “DRILL: Training World Models to Improve Policies, Not Predict Pixels” 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
Zhaolu Kang, Tailong Luo, Lei Wei, Guangyuan Dong, Sihan Liu, Yingjie He, Siheng Wang, Shuaibo Li and 4 more
DRILL trains a policy-aware video world model for robot manipulation by updating the model according to whether imagined rollouts improve policy return, rather than optimizing next-frame pixel error alone. In simulation across five manipulation benchmark suites, DRILL-IMG raises mean policy success by 13.1 percentage points over WMPO, despite 17% higher one-step pixel MSE. These results do not establish real-robot performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.