Explicit Language Memory Enhances Long-Horizon Planning in VLA Models

A recent study available on arXiv (2608.04765) introduces a hierarchical framework for vision-language-action (VLA) models, featuring a dedicated language-memory module to tackle issues in long-horizon robotic tasks. The research highlights four key shortcomings of current VLA models: sparse expert demonstrations impede cross-task compositional generalization; the non-Markovian characteristics of long-horizon tasks hinder policies that rely solely on present observations from maintaining temporal consistency; insufficient closed-loop error correction leads to the accumulation of execution errors; and end-to-end action fine-tuning may degrade the high-level semantic representations of vision-language model (VLM) backbones. The proposed method transforms discrete temporal observations into a cohesive textual memory sequence with temporal logic, separating the system into high-level VLM and low-level action components. This strategy seeks to enhance temporal consistency and error correction by utilizing explicit language memory. The paper is classified as a cross submission and can be accessed via the provided URL.