README.md

metadata

task_categories:
  - robotics
tags:
  - manipulation
  - imitation-learning
  - visuomotor-policies
  - short-term-memory

ReMemBench: Scaling Short-Term Memory of Visuomotor Policies for Long-Horizon Tasks

[Project Page]   [Paper]   [GitHub]

ReMemBench is a benchmark consisting of eight diverse household manipulation tasks spanning four categories of short-term memory, designed to foster general memory mechanisms in robotic visuomotor policies. Built upon RoboCasa, it provides expert teleoperated demonstrations for training and evaluating policies in long-horizon tasks.

Task Categories

Tasks are organized by memory type. Each task is provided with 50 expert demonstrations for training.

Task Name	Memory Category	Task Variants
Retrieve Fruit Remember fruit location (out of view).	Spatial Memory Recall object locations	MemFruitInSinkLeftFar MemFruitInSinkRightFar
Retrieve Oil Remember oil bottle location among distractors.	Spatial Memory Recall object locations	MemRetrieveOilsFromCounterLL MemRetrieveOilsFromCounterLR MemRetrieveOilsFromCounterRL MemRetrieveOilsFromCounterRR
Cook Meat Remember cooking duration while waiting.	Prospective Memory Retain intentions over delay	MemHeatPot
Cook Meat and Vegetable Remember multiple timed actions.	Prospective Memory Retain intentions over delay	MemHeatPotMultiple
Wash and Return to Container Remember which saucer (left/right) the fruit came from.	Object-Associative Memory Recall associations	MemWashAndReturnLeft MemWashAndReturnRight
Wash and Return to Original Spot Remember original countertop location.	Object-Associative Memory Recall associations	MemWashAndReturnSameLocation
Microwave Breadsticks Remember count of breadsticks moved.	Object-Set Memory Maintain/update sets	MemPutKBreadInMicrowave
Relocate Bowls Remember count of bowls among distractors.	Object-Set Memory Maintain/update sets	MemPutKBowlInCabinet

Data Downloading

You can download the dataset using the huggingface-cli:

huggingface-cli download Rutav/ReMemBench-Dataset \
  --repo-type dataset \
  --local-dir ReMemBench-Dataset \
  --local-dir-use-symlinks False

Dataset Structure

The dataset is organized by task name, with each task containing demonstration sessions in HDF5 format.

File Structure

ReMemBench-Dataset/
├── MemFruitInSinkLeftFar/
├── MemHeatPot/
│   ├── [timestamp]/
│   │   ├── demo.hdf5
│   │   └── demo_im128.hdf5  # Image version
├── ...
└── task_embeds_clip_v3.pickle

HDF5 File Structure (demo_im128.hdf5)

• actions: (T, 12) - [7D arm, 4D base, 1D mode]
• obs:
  • robot0_joint_pos_cos / robot0_joint_pos_sin: (T, 7) - Joint position encoding
  • robot0_gripper_qpos: (T, 2) - Gripper position
  • robot0_agentview_center_image: (T, 128, 128, 3) - RGB third-person view
  • robot0_eye_in_hand_image: (T, 128, 128, 3) - RGB eye-in-hand view
• rewards, dones, states: Standard simulation signals

Exploring the Data

To visualize demonstrations, use the replay_dataset.py script from the official repository:

python robocasa/scripts/replay_dataset.py \
  --hdf5_path ReMemBench-Dataset/MemHeatPot/[timestamp]/demo_im128.hdf5 \
  --episode_idx 0 \
  --render

Converting to LeRobot Dataset Format

A conversion script is provided in the repository to port the data to LeRobot format:

python robocasa/scripts/port_to_lerobot.py \
  --dataset_path ReMemBench-Dataset/MemHeatPot/[timestamp]/demo_im128.hdf5 \
  --repo_name your_hf_username/MemHeatPot

Citation

@article{shah2024scaling,
  title={Scaling Short-Term Memory of Visuomotor Policies for Long-Horizon Tasks},
  author={Shah, Rutav and others},
  journal={arXiv preprint arXiv:2606.16178},
  year={2024}
}