Memory-T-Bench

Partially observable, contact-rich 2D pushing tasks derived from Push-T, introduced in "Remember what you did?: Learning Behavioral Memories for Partially Observable Object Manipulation" (paper · project page). The current observation never reveals what to do next; the policy must remember what it already did or found out.

The release ships only the modalities a real-world policy has access to: image, robot state and action. No privileged hidden-state arrays are included, so a model cannot read the variable it is supposed to remember.

Tasks

Multi-Goals

Swap-Direct

Swap-Shuffle

Find-Track

folder	task	what must be remembered	success criterion	episodes	steps
multi-goals	Multi-Goals — Push the gray T block onto each of the three target areas, each of them only once, in any order.	which targets are already done	each goal reached at ≥ 85 % block–goal overlap, none twice	320	94,636
swap-direct	Swap-Direct — Push the blue and red T blocks to the middle, stop on the black dot, then move each block into the other's starting area.	where each T block started	both blocks ≥ 85 % overlap with the other's start area	320	105,615
swap-shuffle	Swap-Shuffle — Use the empty target area as a buffer to shuffle the blue and red T blocks into each other's starting area.	where each T block started	both blocks ≥ 85 % overlap with the other's start area	320	118,629
find-track	Find-Track — Push the gray T block from the starting area through one of three tracks to the target area on the opposite side; two tracks are impassable.	which track is passable	≥ 85 % overlap with the target area, no re-tried track	320	91,807

Each folder holds data.zarr.zip and a preview.gif. Environments, data collection scripts and the CAMP training code are in the CAMP repository (memory_t_bench package).

Quick start

from huggingface_hub import hf_hub_download
import zarr

path = hf_hub_download(repo_id="harrywang01/Memory-T-Bench",
                       filename="multi-goals/data.zarr.zip", repo_type="dataset")
root = zarr.open(zarr.ZipStore(path, mode="r"), mode="r")
print(root.tree())
ends = root["meta/episode_ends"][:]        # exclusive end index of every episode
ep0 = slice(0, int(ends[0]))
img, state, action = root["data/img"][ep0], root["data/state"][ep0], root["data/action"][ep0]

Data format

Each task is a single data.zarr.zip in the Diffusion Policy replay-buffer convention:

data.zarr.zip
├── data/
│   ├── action                (N, 2)             float32   target agent position (pixels)
│   ├── img                   (N, 96, 96, 3)     uint8     rendered observation
│   ├── state                 (N, 2)             float32   agent position (pixels)
│   └── language_instruction  (N,)               <U200     constant per task
└── meta/
    └── episode_ends          (E,)               int64     exclusive end indices

Arrays are flat over all episodes; episode i spans [episode_ends[i-1], episode_ends[i]).

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 — see LICENSE.