jimwang418/MeMoREval-scenes

MeMoREval Scenes

This dataset holds the four simulated indoor environments of MeMoREval, a benchmark for evaluating memory modality in embodied agents. The task specifications and the simulated Dexmate Vega 1 robot model are in the GitHub repository.

• Paper: What Should Embodied Agents Remember? Benchmarking and Evaluating Memory Modality for Embodied Agents
• Project page: https://memoreval.github.io/
• Tasks, robot model and code: https://github.com/jimwang418/MeMoREval

hf download jimwang418/MeMoREval-scenes --repo-type dataset --local-dir data/scenes

Environments

The environments are adapted from SceneSmith scenes and prepared for the Genesis simulator.

Scene	Regions
office	reception, open_office, hallway, meeting_room, meeting_room_2, bathroom, lounge, kitchen
studio	shooting_area, client_lounge, editing_room, dressing_room, props_storage, office, bathroom
home	hallway, living_room, dining_room, kitchen, master_bedroom, kids_bedroom_1, kids_bedroom_2, bathroom_1, bathroom_2
clinic	reception, waiting_room, corridor, vestibule, bathroom_1, bathroom_2, pharmacy, lab, exam_room_1, exam_room_2, exam_room_3, exam_room_4

Region names are the room identifiers used in the task files, for example in navigate_to calls and reference answers.

<scene>/
  genesis_scene.json        Genesis scene description; asset paths are relative to <scene>/
  combined_house/           Drake model directives for the whole scene (*.dmd.yaml)
  room_<name>/              per-room furniture, wall-mounted, ceiling-mounted and manipulable objects
  room_geometry/            per-room geometry
  floor_plans/              floors, walls and windows
  articulated_objects/      articulated furniture as MJCF, with meshes
  materials/                textures
  navigation/
    cspace/                 robot configuration-space maps for navigation planning
    floor_map/              2D floor, obstacle and semantic maps
  package.xml               package manifest from the SceneSmith scene generator

Each object has a visual mesh and a set of convex collision meshes. Articulated furniture such as cabinets, drawers and wardrobes was converted from the SceneSmith assets into MJCF.

Loading

The benchmark code that loads these scenes will be released with the rest of MeMoREval. Meanwhile, <scene>/combined_house/house_furniture_welded.dmd.yaml can be loaded with Drake's model directives. <scene>/genesis_scene.json lists every room frame and the model files placed in it.

License

These scenes are released under the Apache License 2.0. They are adapted from scenes generated with SceneSmith, and each package.xml keeps the license declared by SceneSmith's scene generator. See NOTICE for attributions.

Citation

@misc{wang2026memoreval,
  title         = {What Should Embodied Agents Remember? Benchmarking and Evaluating Memory Modality for Embodied Agents},
  author        = {Wang, Junxiang and Liu, Yuchen and Wu, Yi and Chen, Guofei and Pechuk, Nir and Chen, Tao and Erickson, Zackory and Qin, Yuzhe and Li, Tianyu},
  year          = {2026},
  eprint        = {2610.12561},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2610.12561}
}

Please also cite SceneSmith:

@inproceedings{pfaff2026scenesmith,
  title     = {SceneSmith: Agentic Generation of Simulation-Ready Indoor Scenes},
  author    = {Pfaff, Nicholas and Cohn, Thomas and Zakharov, Sergey and Cory, Rick and Tedrake, Russ},
  booktitle = {International Conference on Machine Learning (ICML)},
  year      = {2026}
}