Memory-Manip-Bench

Memory-dependent tabletop manipulation tasks for a Franka Panda (robosuite 1.5 / MuJoCo 3), introduced in "Remember what you did?: Learning Behavioral Memories for Partially Observable Object Manipulation" (paper · project page). Every task hides one piece of information that the policy can only discover by interacting and must then remember; forgetting it makes the episode fail.

Tasks

Uncover-Blocks	Swap-Block	Probe-Insert	Button-Lightbulb
Stack-Lego	Find-Soda	Open-Door

file	task	env (env_name)	task camera	what must be remembered
image_uncoverblock.hdf5	Uncover-Blocks	UncoverBlockVisionHard	agentview	which covers were already lifted
image_blockswap.hdf5	Swap-Block	FruitSwapEasy	peg_focus_view	where each cube started
image_insertpeg.hdf5	Probe-Insert	InsertPegVision	peg_focus_view	which holes were already tried
image_buttonlightbulb.hdf5	Button-Lightbulb	ButtonLightbulbVision	agentview	which button toggles which bulb
image_legostacking.hdf5	Stack-Lego	LegoStackingVision	peg_focus_view	which colour is the solid-bottom brick
image_findsoda.hdf5	Find-Soda	FindSodaVision	drawerview	which drawers were already checked
image_opendoors.hdf5	Open-Door	OpenDoorsVision	doorview	which doors were already tried

One robomimic-style hdf5 of teleoperated demonstrations per task (100 demos each).

Environments, teleoperation scripts and the CAMP training code are in the CAMP repository (memory_manip_bench package).

Download

pip install -U "huggingface_hub[cli]"
huggingface-cli download harrywang01/Memory-Manip-Bench --repo-type dataset \
    --include image_uncoverblock.hdf5 --local-dir data/

Data format

data/                              (attrs: env_args — env_name + env_kwargs to rebuild the env)
└── demo_<i>/                      (attrs: num_samples, model_file; Find-Soda also ep_meta)
    ├── obs/<task camera>_image    (T, 256, 256, 3) uint8
    ├── obs/robot0_eye_in_hand_image (T, 256, 256, 3) uint8
    ├── obs/robot0_eef_pos         (T, 3)  float
    ├── obs/robot0_eef_quat        (T, 4)  float
    ├── obs/robot0_gripper_qpos    (T, 2)  float
    ├── actions                    (T, 7)  delta OSC command
    ├── actions_abs                (T, 7)  absolute [goal_pos(3), goal_axis_angle(3), gripper(1)]
    ├── rewards, dones             (T,)
    └── states                     (T, D)  MuJoCo states

import h5py, json
with h5py.File("data/image_uncoverblock.hdf5", "r") as f:
    env_args = json.loads(f["data"].attrs["env_args"])
    demo = f["data/demo_1"]
    imgs, acts = demo["obs/agentview_image"][:], demo["actions_abs"][:]

model_file (the demo's MuJoCo model) and states replay a demo in the environment. Find-Soda's placement target is a model parameter, not simulator state, so its demos also carry ep_meta = {"target_xy", "soda_drawer_idx"}, which FindSodaVision.set_ep_meta restores on replay.

Changelog

• 2026-10-11, image_findsoda.hdf5: each demo's model_file now places the target patch where the images show it (under the placed can), and ep_meta was added. Before, model_file held an unrelated patch position, so replaying a demo from its states showed the can next to the patch. Images, states and actions are unchanged. media/find_soda.gif was re-rendered.

Citation

@misc{wang2026rememberdidlearningbehavioral,
      title={Remember what you did?: Learning Behavioral Memories for Partially Observable Object Manipulation},
      author={Kuancheng Wang and Seungho Yeom and Jinglin Cao and Yuheng Zhi and Nikhil Shinde and Michael Yip},
      year={2026},
      eprint={2606.21188},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2606.21188},
}

License

MIT.