VPWEM: Non-Markovian Visuomotor Policy with Working and Episodic Memory

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Lei, Yuheng, Liang, Zhixuan, Zhang, Hongyuan, Luo, Ping
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912944383590400
author Lei, Yuheng
Liang, Zhixuan
Zhang, Hongyuan
Luo, Ping
author_facet Lei, Yuheng
Liang, Zhixuan
Zhang, Hongyuan
Luo, Ping
contents Imitation learning from human demonstrations has achieved significant success in robotic control, yet most visuomotor policies still condition on single-step observations or short-context histories, making them struggle with non-Markovian tasks that require long-term memory. Simply enlarging the context window incurs substantial computational and memory costs and encourages overfitting to spurious correlations, leading to catastrophic failures under distribution shift and violating real-time constraints in robotic systems. By contrast, humans can compress important past experiences into long-term memories and exploit them to solve tasks throughout their lifetime. In this paper, we propose VPWEM, a non-Markovian visuomotor policy equipped with working and episodic memories. VPWEM retains a sliding window of recent observation tokens as short-term working memory, and introduces a Transformer-based contextual memory compressor that recursively converts out-of-window observations into a fixed number of episodic memory tokens. The compressor uses self-attention over a cache of past summary tokens and cross-attention over a cache of historical observations, and is trained jointly with the policy. We instantiate VPWEM on diffusion policies to exploit both short-term and episode-wide information for action generation with nearly constant memory and computation per step. Experiments demonstrate that VPWEM outperforms state-of-the-art baselines including diffusion policies and vision-language-action (VLA) models by more than 20% on the memory-intensive manipulation tasks in MIKASA and achieves an average 5% improvement on the mobile manipulation benchmark MoMaRT. Code is available at https://github.com/HarryLui98/code_vpwem.
format Preprint
id arxiv_https___arxiv_org_abs_2603_04910
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle VPWEM: Non-Markovian Visuomotor Policy with Working and Episodic Memory
Lei, Yuheng
Liang, Zhixuan
Zhang, Hongyuan
Luo, Ping
Robotics
Artificial Intelligence
Machine Learning
Imitation learning from human demonstrations has achieved significant success in robotic control, yet most visuomotor policies still condition on single-step observations or short-context histories, making them struggle with non-Markovian tasks that require long-term memory. Simply enlarging the context window incurs substantial computational and memory costs and encourages overfitting to spurious correlations, leading to catastrophic failures under distribution shift and violating real-time constraints in robotic systems. By contrast, humans can compress important past experiences into long-term memories and exploit them to solve tasks throughout their lifetime. In this paper, we propose VPWEM, a non-Markovian visuomotor policy equipped with working and episodic memories. VPWEM retains a sliding window of recent observation tokens as short-term working memory, and introduces a Transformer-based contextual memory compressor that recursively converts out-of-window observations into a fixed number of episodic memory tokens. The compressor uses self-attention over a cache of past summary tokens and cross-attention over a cache of historical observations, and is trained jointly with the policy. We instantiate VPWEM on diffusion policies to exploit both short-term and episode-wide information for action generation with nearly constant memory and computation per step. Experiments demonstrate that VPWEM outperforms state-of-the-art baselines including diffusion policies and vision-language-action (VLA) models by more than 20% on the memory-intensive manipulation tasks in MIKASA and achieves an average 5% improvement on the mobile manipulation benchmark MoMaRT. Code is available at https://github.com/HarryLui98/code_vpwem.
title VPWEM: Non-Markovian Visuomotor Policy with Working and Episodic Memory
topic Robotics
Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2603.04910