TPRVLA:Temporal-Preference-Recovery Vision-Language-Action Model

Vision-language-action (VLA) models offer a promising approach to robotic manipulation, yet reliable long-horizon execution in dynamic environments remains challenging. OpenVLA supports strong task generalization and parameter-efficient adaptation, but its single-image input limits temporal modeling, while its standard framework lacks explicit use of failure experience and runtime recovery mechanisms. To address these limitations, we propose TPRVLA, which integrates temporal memory, preference learning, and closed-loop recovery. A lightweight temporal memory adapter incorporates consecutive observations into the frozen OpenVLA backbone to capture short-term dynamics. Weighted preference pairs are constructed from successful and failed trajectories under similar visual observations and identical instructions. The policy is then adapted through a joint direct preference optimization (DPO) and behavior cloning (BC) objective to favor preferred actions. During execution, a rule-based state monitor combines end-effector motion, action repetition, and visual changes to trigger checkpoint-guided recovery when anomalies persist. Evaluations across 900 rollouts on ten LIBERO Spatial tasks with three random seeds show improved task performance, particularly on tasks requiring extended execution and closed-loop adjustment, while increasing peak GPU memory usage by less than 1%.