nielsr's picture
nielsr HF Staff
Improve dataset card: add robotics metadata, links, and sample usage
d982c4e verified
|
Raw
History Blame
2.12 kB
metadata
license: apache-2.0
task_categories:
  - robotics

RoboMME Training Data (Pickle Format)

Paper | Website | Benchmark Code | Policy Learning Code

This repository contains preprocessed pickle files for RoboMME training data and npy files for cached image tokens. This dataset is used in the MME-VLA experiments.

RoboMME is a large-scale standardized benchmark for evaluating and advancing Vision-Language-Action (VLA) models in long-horizon, history-dependent scenarios. It comprises 16 manipulation tasks across four cognitively motivated suites:

  • Counting (Temporal memory)
  • Permanence (Spatial memory)
  • Reference (Object memory)
  • Imitation (Procedural memory)

Repository Structure

.
├── data        # zipped pickle files
├── features    # zipped precompute siglip embeddings
├── meta        # statistics for robomme
├── memer       # VLM subgoal training data for MemER (only used for symbolic memory)
├── qwenvl      # VLM subgoal training data for QwenVL  (only used for symbolic memory)
└── README.md

Sample Usage

To evaluate on the test set using the BenchmarkEnvBuilder from the benchmark repository:

task_id = "PickXtimes"
episode_idx = 0
env_builder = BenchmarkEnvBuilder(
    env_id=task_id,
    dataset="test",
)

env = env_builder.make_env_for_episode(episode_idx)
obs, info = env.reset() # initial step
task_goal = info['task_goal'][0]

# Policy interaction loop
# obs, _, terminated, truncated, info = env.step(action)

Citation

@article{dai2026robomme,
  title={RoboMME: Benchmarking and Understanding Memory for Robotic Generalist Policies},
  author={Dai, Yinpei and Fu, Hongze and Lee, Jayjun and Liu, Yuejiang and Zhang, Haoran and Yang, Jianing and Chelsea Finn and Nima Fazeli and Joyce Chai},
  journal={arXiv preprint arXiv:2603.04639},
  year={2026}
}