Reproduce “EmbodiedMind: Adaptive Data Curation and Prefix-Tree Reinforcement Learning for Efficient Embodied Intelligence” 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
Feifan Wang, Zongbing Zhang, Yu Zhang, Lingfeng Wang, Yurui Zhu, Jin Deng, Mingliang Zhang, Zhengguang Gao and 3 more
EmbodiedMind trains embodied models by filtering and balancing examples, then uses Trie-GRPO to compare next actions sharing an action prefix, improving credit assignment in long plans. Its 18-benchmark average is 70.02% for EmbodiedMind-8B versus 67.98% for RoboBrain2.5-8B, using 522K versus 12.4M samples. On VLM-PlanSim-99, Trie-GRPO scores 66.33 ± 0.82 versus 61.50 ± 0.84 for GRPO across five seeds; these comparisons do not establish universal or causal effects.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.