README.md

metadata

license: mit
task_categories:
  - robotics

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

CVPR 2026

Paper | Project Page | GitHub

LMEE-Bench is a benchmark for Long-term Memory Embodied Exploration (LMEE). It is designed to evaluate an agent's exploratory cognition and decision-making behaviors by incorporating multi-goal navigation and memory-based question answering tasks.

Dataset Structure

The benchmark consists of the following components:

• lmee_bench_sub: Includes 58 tasks.
• lmee_bench: Includes the full 166 tasks.
• task_test: Trajectory data test set.

Sample Usage (Evaluation)

To evaluate a model on LMEE-Bench, you can follow the instructions provided in the official repository.

1. 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.

2. Evaluation

After running the reasoning script, you will get a results file (e.g., lmee_answer.json). 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"

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}
}