README.md

metadata

license: apache-2.0
task_categories:
  - robotics

RoboMemArena: A Comprehensive and Challenging Robotic Memory Benchmark

Project Page | Paper | GitHub

RoboMemArena is a large-scale benchmark of 26 tasks designed to evaluate robotic memory. It features average trajectory lengths exceeding 1,000 steps per task, with 68.9% of subtasks being memory-dependent. The benchmark covers various domains including multi-sequence execution, occlusion handling, counting, and object transferring.

Task Categories

This dataset contains four categories of tasks:

Category	Description	Task IDs
Multi-Sequence	Tasks requiring execution of multiple sequential sub-steps	Task1, Task2, Task3, Task22
Multi-Occlusion	Tasks involving occluded or hidden objects	Task4, Task5, Task11, Task12, Task13, Task14, Task17, Task20, Task21, Task23, Task24
Multi-Counting	Tasks requiring repeated actions, such as pouring twice	Task6, Task7, Task8, Task9, Task10, Task15, Task16
Multi-Transferring	Tasks involving transferring objects between locations	Task18, Task19, Task25, Task26

Dataset Structure

The dataset is organized into category folders, each containing task subfolders. The demonstration data is provided in HDF5 format with keyframe annotations.

<dataset_root>/
├── <category_1>/
│   └── 1_cookies_tomato_basket_dataset/
│       └── subtask_data/         # Keyframe-annotated HDF5 episodes
│           ├── pick_cookies_0_seed100_task1.hdf5
│           ├── pick_cookies_0_seed101_task1.hdf5
│           └── ...

Each HDF5 file in subtask_data/ contains:

• data/demo_{id}/actions: (T, 7) end-effector actions
• data/demo_{id}/obs/agentview_rgb: (T, 256, 256, 3) top-down view
• data/demo_{id}/obs/eye_in_hand_rgb: (T, 256, 256, 3) wrist camera
• data/demo_{id}/obs/ee_states, gripper_states, joint_states: Robot state

Sample Usage (RLDS Conversion)

You can use the RoboMemArena_dataset_builder.py script provided in the GitHub repository to convert the HDF5 data to RLDS (TFDS) format:

import RoboMemArena_dataset_builder as b
import tensorflow_datasets as tfds

# Set the source dataset root and run the builder
ds_builder = b.RoboMemArenaDataset(data_dir='/path/to/output')
ds_builder.download_and_prepare()

Citation

@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}
}