MVP: Memory-enhanced Vision-Language-Action Policy with Feedback Learning

ICLR 2026 Conference Submission11853 Authors

Keywords: Robotic Manipulation, Vision-Language-Action Models, Memory

TL;DR: A memory-enhanced Vision-Language-Action model

Abstract: Recent advances in Vision-Language-Action (VLA) models have enabled robots to perform a wide range of manipulation tasks conditioned on language instructions, offering strong generalization across tasks, objects, and environments. However, most existing VLAs operate under a Markov assumption, limiting their ability to handle temporally extended tasks and learn from feedback. To address these limitations, we propose MVP, a non-Markovian VLA model that leverages episodic memory composed of historical actions and visual observations. To mitigate the computational cost of storing high-dimensional histories, we introduce a compact memory representation inspired by video understanding techniques. Additionally, to prevent the model from disregarding historical inputs during training, we design a novel feedback learning strategy based on SO(3) trajectory perturbation. This approach encourages the model to associate actions with their environmental consequences through observation-action-observation sequences. Experimental results on both simulated and real-world benchmarks demonstrate that MVP outperforms existing models, particularly on tasks that require temporal reasoning and history-dependent decision-making. Our findings highlight the importance of memory and feedback in advancing the capabilities of general-purpose robotic manipulation systems.

Primary Area: applications to robotics, autonomy, planning

Submission Number: 11853