Evomem vla improves long task memory for robot manipulation
EvoMem-VLA: State-Evolution Memory for Long-Horizon Robot Manipulation
RoboticsComputer Vision and Pattern Recognition
Summary
Robots sometimes struggle to remember important information during long tasks when objects or changes they need to track go out of view. The authors present EvoMem-VLA, a way for robots to remember how things changed over time, not just snapshots of past scenes. This approach helps robots better understand what happened during multi-step tasks, significantly improving their performance in both simulated and real-world tests. EvoMem-VLA uses smart ways to encode differences between frames and adapts for tasks with multiple stages, enabling robots to plan actions more reliably.
What this means in practice
- •For robotics engineers: Improve robot manipulation in complex tasks by retaining and using detailed change information between observations for better action decisions.
- •For automation system integrators: Enable deployment of service robots that reliably perform long and multi-step tasks by incorporating enhanced memory of past interactions.$Commercial implications: This approach supports development of more capable service robots that can be sold for automation in manufacturing and logistics.
Authors
Yuheng Na, Zhide Zhong, Junjie He, Junfeng Li, Haodong Yan, Jiaan Wang, Jiaguan Zhu, Yangyang Zheng, Tianyu Huang, Haoang Li
Abstract
Most vision-language-action (VLA) models rely on current observations and lose task-relevant evidence once it leaves view, limiting performance on long-horizon, memory-dependent tasks. Existing efforts incorporate compressed historical features or sparse visual keyframes. However, isolated snapshots can leave the policy uncertain about what changed during past interactions and which action should follow. To overcome this limitation, we propose EvoMem-VLA, which constructs state-evolution memory by explicitly encoding and retaining observed changes between historical states. These change representations preserve evidence of interaction outcomes, allowing the policy to track task progress beyond isolated snapshots. Specifically, we introduce conditional delta tokenization to encode ordered frame pairs into directional, source-conditioned delta tokens, each associated with its corresponding state evidence. A shared VLM backbone supports task-adaptive routing: normal long-horizon tasks follow a direct action route, whereas multi-stage tasks use a subtask route that generates an executable subtask as an additional input for action generation. With a single jointly trained policy for each simulation benchmark, EvoMem-VLA achieves success rates of 80.7\% on RMBench, 82.0\% on RoboMME and 83.8\% across four real-world tasks spanning two robot embodiments. These results represent substantial improvements over the previous state of the art in all three evaluation settings.