arxivcs.RO2026-07-09
TFP: Temporally Conditioned Memory-Fusion Policies for Visuomotor Learning
Yushen Liang, Yue Peng, Baosheng Jin, Tianluo Zhang, Xinyu Zhang, Shuyi Zhou, et al.
Vision--Language--Action (VLA) policies such as $π_{0.5}$ and OpenVLA perform well on many manipulation tasks, but they are often reactive: the next action is predicted from the current observation, instruction, and proprioceptive state. This assumption breaks down in stage-depen…