[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-79943":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":9,"language":10,"languages":9,"totalLinesOfCode":9,"stars":11,"forks":12,"watchers":13,"openIssues":12,"contributorsCount":14,"subscribersCount":14,"size":14,"stars1d":14,"stars7d":12,"stars30d":15,"stars90d":14,"forks30d":14,"starsTrendScore":16,"compositeScore":17,"rankGlobal":9,"rankLanguage":9,"license":9,"archived":18,"fork":18,"defaultBranch":19,"hasWiki":18,"hasPages":18,"topics":20,"createdAt":9,"pushedAt":9,"updatedAt":21,"readmeContent":22,"aiSummary":23,"trendingCount":14,"starSnapshotCount":14,"syncStatus":16,"lastSyncTime":24,"discoverSource":25},79943,"RoboMemArena","OpenHelix-Team\u002FRoboMemArena","OpenHelix-Team","RoboMemArena: A Comprehensive and Challenging Robotic Memory Benchmark",null,"Python",92,4,1,0,11,2,42.2,false,"main",[],"2026-06-12 04:01:25","# RoboMemArena: A Comprehensive and Challenging Robotic Memory Benchmark\n\nRoboMemArena is a comprehensive and challenging robotic memory benchmark with 26 manipulation tasks, demonstration data, and evaluation BDDL.\n\n## Links\n\n[![arXiv](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2605.10921-b31b1b?style=for-the-badge)](https:\u002F\u002Farxiv.org\u002Fhtml\u002F2605.10921v1)\n[![Project Page](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FProject-Page-76b900?style=for-the-badge)](https:\u002F\u002Frobomemarena.github.io\u002F)\n[![Leaderboard Results](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLeaderboard-Results-007ec6?style=for-the-badge)](https:\u002F\u002Frobomemarena.github.io\u002Fleaderboard.html)\n[![Dataset Hugging Face](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FDataset-Hugging%20Face-f3b900?style=for-the-badge)](https:\u002F\u002Fhuggingface.co\u002Fdatasets\u002FRoboMemArenaBenchmark\u002FRoboMemArena)\n\n## Dataset Structure\n\nThe dataset is hosted on Hugging Face:\n\n[![Dataset Hugging Face](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FDataset-Hugging%20Face-f3b900?style=for-the-badge)](https:\u002F\u002Fhuggingface.co\u002Fdatasets\u002FRoboMemArenaBenchmark\u002FRoboMemArena)\n\nThe released dataset is organized as:\n\n- 4 high-level category folders\n- each category folder contains task subfolders (for example, `1_cookies_tomato_basket_dataset`)\n- each task subfolder keeps the same internal structure\n\n```\n\u003Cdataset_root>\u002F\n├── \u003Ccategory_1>\u002F\n│   └── 1_cookies_tomato_basket_dataset\u002F\n│       └── subtask_data\u002F         # Keyframe-annotated HDF5 episodes (used for training)\n│           ├── pick_cookies_0_seed100_task1.hdf5\n│           ├── pick_cookies_0_seed101_task1.hdf5\n│           └── ...\n├── \u003Ccategory_2>\u002F\n├── \u003Ccategory_3>\u002F\n└── \u003Ccategory_4>\u002F\n```\n\nFilename convention for subtask HDF5 files:\n`\u003Cprimitive>_\u003Csubtask_order>_seed\u003Cseed>_task\u003Ctask_id>.hdf5`, where\n`subtask_order` is the 0-based order index used in task decomposition.\n\nThe same task seed links multiple subtask files into one complete long-horizon\ntrajectory. For example, files with `seed100` under the same task folder should\nbe sorted by `subtask_order` and concatenated to reconstruct the full task\nepisode. If your training target is the complete task rather than individual\nsubtasks, group files by `task_id` and `seed`, then concatenate the ordered\nsubtask episodes before training.\n\n### Key Directories\n\n| Directory | Description |\n|-----------|-------------|\n| **subtask_data\u002F** | Sub-episodes with **keyframe annotations**; each HDF5 contains `data\u002Fdemo_*` with `actions`, `obs\u002Fagentview_rgb`, `obs\u002Feye_in_hand_rgb`, `obs\u002Fee_states`, `obs\u002Fgripper_states`, `obs\u002Fjoint_states` |\n\n### HDF5 Format (subtask_data)\n\n- `data\u002Fdemo_{id}\u002Factions` — (T, 7) end-effector actions\n- `data\u002Fdemo_{id}\u002Fobs\u002Fagentview_rgb` — (T, 256, 256, 3) top-down view\n- `data\u002Fdemo_{id}\u002Fobs\u002Feye_in_hand_rgb` — (T, 256, 256, 3) wrist camera\n- `data\u002Fdemo_{id}\u002Fobs\u002Fee_states`, `gripper_states`, `joint_states` — robot state\n\n## RLDS Conversion\n\nUse `RoboMemArena_dataset_builder.py` to convert HDF5 to RLDS (TFDS) format:\n\nRLDS conversion is a data-only step; a CPU Python environment is sufficient.\n\n- Python 3.9\n- `tensorflow==2.13.0`\n- `tensorflow-datasets==4.9.2`\n- `h5py==3.9.0`\n- `numpy==1.24.3`\n\nExample environment setup:\n\n```bash\nconda create -n robomemarena-rlds python=3.9 -y\nconda activate robomemarena-rlds\npip install tensorflow==2.13.0 tensorflow-datasets==4.9.2 h5py==3.9.0 numpy==1.24.3\n```\n\nSet the source dataset root and run the builder from this repository root:\n\n```bash\nexport ROBOMEMARENA_DATA_ROOT=\u002Fpath\u002Fto\u002F\u003Cdataset_root>\npython -c \"\nimport RoboMemArena_dataset_builder as b\nimport tensorflow_datasets as tfds\nds_builder = b.RoboMemArenaDataset(data_dir='\u002Fpath\u002Fto\u002Foutput')\nds_builder.download_and_prepare()\n\"\n```\n\n## Evaluation BDDL\n\nEvaluation tasks are defined in the `bddl\u002F` folder:\n\n- `1_*.bddl` … `26_*.bddl` — 26 BDDL task definitions using contiguous benchmark task IDs\n\nThese BDDL files can be used with the provided LIBERO-compatible evaluation environment.\nThe full 26-task benchmark descriptions are available here:\n\n- [Benchmark Task Details](https:\u002F\u002Frobomemarena.github.io\u002F#task-details)\n- [Evaluation Benchmark Overview](evaluation_benchmark\u002FREADME.md)\n- [26-Task Reference Evaluation](evaluation_benchmark\u002Freference_evaluation\u002FREADME.md)\n- [Task Evaluation Code Guide](evaluation_benchmark\u002Fdocs\u002Ftask_evaluation_code_guide.md)\n\n## OpenPI Runtime Interface\n\nEvaluation scripts support two OpenPI runtime paths:\n\n- Bundled minimal runtime (default): `third_party\u002Fopenpi_minimal`\n- External OpenPI source (optional): set `OPENPI_ROOT=\u002Fabs\u002Fpath\u002Fto\u002Fopenpi`\n\nIf using external OpenPI source, ensure the following exist:\n\n- `scripts\u002Fserve_policy.py`\n- `packages\u002Fopenpi\u002Fsrc`\n- `packages\u002Fopenpi-client\u002Fsrc`\n\n## LIBERO Environment Setup\n\nEvaluation requires a LIBERO-compatible environment. This repo includes a local fork under `evaluation_benchmark\u002Flibero_fork\u002F`:\n\n```bash\n# Clone RoboMemArena (includes submodules)\ngit clone --recurse-submodules https:\u002F\u002Fgithub.com\u002FOpenHelix-Team\u002FRoboMemArena.git\ncd RoboMemArena\n\n# If you already cloned without --recurse-submodules:\n# git submodule update --init --recursive\n\n# Make the local LIBERO fork importable\nexport PYTHONPATH=\"${PWD}\u002Fevaluation_benchmark\u002Flibero_fork:${PYTHONPATH}\"\n```\n\nCopy the `bddl\u002F` folder from this repo to your evaluation config path, or set `LIBERO_BDDL_PATH` to point to `bddl\u002F` for RoboMemArena task definitions.\n\n## PrediMem S2 Training Add-on\n\nThis repo also includes a minimal add-on folder:\n\n- [predictive_coding_head\u002F](predictive_coding_head\u002F)\n\nIt provides reusable integration logic for PrediMem S2 training with a **Predictive Coding Head** in a Qwen3-VL style training pipeline.\nSee:\n\n- [predictive_coding_head\u002FREADME.md](predictive_coding_head\u002FREADME.md)\n\nFor the S1 low-level policy, both the environment setup and the training logic\ncan directly follow the official OpenPI repository:\nhttps:\u002F\u002Fgithub.com\u002FPhysical-Intelligence\u002Fopenpi\n\nFor VLM training data construction, users may follow the official Qwen3-VL\nmultimodal `messages` \u002F `content` format described here:\nhttps:\u002F\u002Fgithub.com\u002FQwenLM\u002FQwen3-VL?tab=readme-ov-file#using-transformers-to-chat\n\n## Contact\n\nFor questions, please contact me via WeChat: `leshuaigeye` or email: `leihuashuohit@gmail.com`.\n\n## Citation\n\nIf you find RoboMemArena useful in your research, please cite:\n\n```bibtex\n@article{robomemarena2025,\n  title   = {RoboMemArena: A Comprehensive and Challenging Robotic Memory Benchmark},\n  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},\n  journal = {arXiv preprint arXiv:2605.10921},\n  year    = {2026}\n}\n```\n","RoboMemArena是一个全面且具有挑战性的机器人记忆基准测试平台，包含26个操作任务、演示数据和评估BDDL。该项目使用Python开发，提供了详细的HDF5格式数据集，包括动作、观察（如顶部视角和手腕相机图像）以及机器人状态信息。这些数据被组织成四个高层次类别文件夹，每个类别下有具体的任务子文件夹，并且支持通过种子值链接多个子任务文件来重建完整的任务轨迹。RoboMemArena特别适合于需要对机器人长期记忆能力和复杂任务处理能力进行评估的研究场景，同时也为基于强化学习的算法训练提供了丰富的资源。","2026-06-11 03:58:39","CREATED_QUERY"]