像世界模型一样思考,像 VLA 一样行动:将世界模型表示蒸馏为紧凑型机器人策略
TLDR
通过特征对齐将冻结世界模型表征蒸馏进紧凑VLA策略,提升LIBERO/RoboCasa基准与真机表现,并保持32 ms推理。
评分理由
The paper presents a practical method to transfer world-model grounding to robot policies without generative inference, with solid experiments across benchmarks and real hardware. Weaknesses: limited visibility into method details and failure cases from the abstract; the exact mechanism by which alignment transfers grounding is only asserted.
Read-first 评分解释
综合优先阅读分 22.7,由主题、引用、图谱、方法、可复现性和近期性等信号加权得到。 原始总分保留为 22。
研究版图角色
前沿论文
排序敏感性
稳定性:volatile;排名波动范围:162。