Reproduce “Grana: Compositional Reward Modeling for Long-Horizon Robot Manipulation” 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
Zi-ang Cao, Kaiyuan Zheng, Yinzhen Xu, Haonan Duan, Jiasheng Gu, Fengyuan Hu, Nan Zhu, Kun Lei and 9 more
Grana estimates robot-task progress by selecting an unfinished subtask and predicting its local progress, then adding credit for preceding completed subtasks. Sparse boundary and action-quality annotations train its shared vision-language predictor to represent setbacks and recovery. On 50 held-out tasks, zero-shot RMSE was 0.180 versus 0.227 for fine-tuned RoboMeter, a 20.7% reduction, but this does not guarantee performance on arbitrary unseen tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.