RoboMemArena: A Comprehensive and Challenging Robotic Memory Benchmark

Benchmark Examples

Overview

Figure 1. RoboMemArena converts natural-language instructions into keyframe-annotated trajectories through VLM-based task decomposition, autonomous execution, closed-loop verification, and targeted human refinement of unsuitable annotations.

Abstract

Memory is a critical component of robotic intelligence, as robots must rely on past observations and actions to accomplish long-horizon tasks in partially observable environments. However, existing robotic memory benchmarks still lack multimodal annotations for memory formation, provide limited task coverage and structural complexity, and remain restricted to simulation without real-world evaluation. We address this gap with RoboMemArena, a large-scale benchmark of 26 tasks, with average trajectory lengths exceeding 1,000 steps per task and 68.9% of subtasks being memory-dependent. Its generation pipeline uses a vision-language model (VLM) to design and compose subtasks, generates full trajectories through atomic functions, and provides memory-related annotations, including subtask instructions and native keyframe annotations, while paired real-world memory tasks support physical evaluation. We further design PrediMem, a dual-system VLA in which a high-level VLM planner manages a memory bank with recent and keyframe buffers and uses a predictive coding head to improve sensitivity to task dynamics. Extensive experiments on RoboMemArena show that PrediMem outperforms all baselines and provides insights into memory management, model architecture, and scaling laws for complex memory systems.

PrediMem

PrediMem is a dual-system memory framework built around a high-level VLM planner and a low-level VLA executor. The high-level planner receives the current observation together with a memory bank composed of a recent-frame buffer and a keyframe buffer. It predicts the current subtask and decides whether the current frame should be stored as a keyframe, while the low-level policy executes the latest subtask-conditioned action chunk.

During training, PrediMem adds a predictive coding head that predicts the next-frame visual latent representation from the current hidden state. This auxiliary objective makes the VLM representation more sensitive to physical state transitions such as grasp closure, object placement, repeated pouring, and drawer closure. The predictive coding head is removed at inference time, so the deployed system keeps the same dual-system architecture and inference cost.

Benchmark Results

BibTeX

@article{robomemarena2025,
  title   = {RoboMemArena: A Comprehensive and Challenging
             Robotic Memory Benchmark},
  author  = {Huashuo Lei and Wenxuan Song and Huarui Zhang and Jieyuan Pei and Jiayi Chen and Haodong Yan and Han Zhao and Pengxiang Ding and Zhipeng Zhang and Lida Huang and Donglin Wang and Yan Wang and Haoang Li},
  journal = {arXiv preprint arXiv:2605.10921},
  year    = {2026}
}

Benchmark Task Details

Overview of all 26 RoboMemArena tasks, their corresponding memory types, average total timesteps, and key challenges.