Explore with Long-term Memory: A Benchmark and Multimodal LLM-based Reinforcement Learning Framework for Embodied Exploration

CVPR 2026

Sen Wang, Bangwei Liu, Zhenkun Gao, Lizhuang Ma, Xuhong Wang, Yuan Xie, Xin Tan

---

This is the official repository of Explore with Long-term Memory: A Benchmark and Multimodal LLM-based Reinforcement Learning Framework for Embodied Exploration.

---

News

• [2026/03] Training code for MemoryExplorer is released.
• [2026/03] Inference code for LMEE-Bench is released.
• [2026/02] Our paper is accepted to CVPR 2026!
• [2026/01] Paper is on arXiv.

Preparations

• (All) Download the train and val split of HM3D-Sem
• (Evaluation) Download LMEE-Bench: LMEE-Bench Dataset
• (Evaluation) Download MemoryExplorer Model: MemoryExplorer
• (Training) Download LMEE: LMEE Training Dataset

Put them into the data folder. The final file format should be:

data
├── LMEE-train
│   ├── task_train
│   │   ├── easy
│   │   ├── hard
│   │   ├── medium
│   ├── train_data.parquet
│   │   
├── LMEE-Bench
│   ├── lmee_bench
│   ├── lmee_bench_sub
│   ├── task_test
│   │   ├── easy
│   │   ├── hard
│   │   ├── medium
│   │   
├── ...

Evaluation

1. Installation

Set up the conda environment (Linux, Python 3.9):

cd evaluation

conda create -n lmee python=3.9 -y && conda activate lmee
pip install torch==2.4.1 torchvision==0.19.1 torchaudio==2.4.1 --index-url https://download.pytorch.org/whl/cu121
conda install https://anaconda.org/pytorch3d/pytorch3d/0.7.8/download/linux-64/pytorch3d-0.7.8-py39_cu121_pyt241.tar.bz2 -y
pip install -r requirements.txt
conda install -c conda-forge -c aihabitat habitat-sim=0.3.1 headless faiss-cpu=1.7.4 -y

2. Reasoning

Specify the paths in the configuration file: cfg/eval_lmee_bench.yaml and execute the following command:

python run_lmee.py -cf cfg/eval_lmee_bench.yaml --answer_type open

• answer_type: Choose between open and choice.
• Subset Options:
  • LMEE-Bench/lmee_bench_sub: Includes 58 tasks.
  • LMEE-Bench/lmee_bench: Includes the full 166 tasks.

3. Evaluation

After running the reasoning script, you will get the results file: lmee_answer.json and use the following command to evaluate the question-answering performance:

python eval_lmee_bench.py --json_path "results/exp_eval_lmee/lmee_answer.json" --root_dir "../data/LMEE-Bench/task_test"

Training

1. Installation

Set up the conda environment (Linux, Python 3.10):

cd train

conda create -n memoryexplorer python=3.10 -y && conda activate memoryexplorer
pip install torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0 --index-url https://download.pytorch.org/whl/cu124
pip install hf_transfer vllm==0.8.5.post1 triton==3.2.0
pip install -e .

2. Running

Specify the paths in the running file train_my.sh and IMAGE_ROOT in verl\tooluse\memory_tool.py, and execute the following command:

bash train_my.sh

Please see EasyR1 and verl for more training details.

Todo List

• Release training scripts and dataset
• Release data generation scripts

Acknowledgement

The codebase is built upon 3D-Mem, MemoryEQA and VTool-R1. We thank the authors for their great work.

Citation

@inproceedings{wang2026explore,
  title={Explore with Long-term Memory: A Benchmark and Multimodal LLM-based Reinforcement Learning Framework for Embodied Exploration},
  author={Wang, Sen and Liu, Bangwei and Gao, Zhenkun and Ma, Lizhuang and Wang, Xuhong and Xie, Yuan and Tan, Xin},
  booktitle={Proceedings of the IEEE/CVF Computer Vision and Pattern Recognition (CVPR)},
  year={2026}
}