Memory-Manip-Bench: 3D Contact-Rich Memory-Based Manipulation Tasks

A benchmark dataset for memory-based robot manipulation in 3D environments. Tasks require the robot to recall and act on observations from earlier in the episode — not solvable by purely reactive policies. Every demonstration is annotated with a natural-language task instruction, so the data can be used directly to train language-conditioned policies / VLAs.

Built on Robosuite with teleoperated expert demonstrations.

---

Tasks

Task	Demos	Steps	Cameras	Language instruction
FruitSwapEasy	100	45,432	agentview, eye_in_hand	swap the position of two blocks and use the empty circle as the buffer
FruitSwapHard	100	57,697	agentview, eye_in_hand	swap the position of two blocks and use the empty circle as the buffer
PegInsertion	100	48,366	agentview, peg_focus_view, eye_in_hand	find the correct hole to fully insert the peg.
UncoverBlock	100	39,302	agentview, eye_in_hand	lift the cover to reveal the hidden red block.

• FruitSwap — pick up two blocks and swap their positions, using the empty circle as a temporary buffer; the robot must remember the initial placement before moving the first object. Easy and Hard differ in episode length / layout difficulty.
• PegInsertion — among several candidate holes, fully insert the peg into the correct one (which must be recalled from earlier in the episode).
• UncoverBlock — lift the cover to reveal a hidden red block.

---

Language Conditioning

Every demo carries the task instruction so it can be loaded alongside observations and actions:

/data/demo_N/language_instruction              (T,)  variable-length UTF-8 string
/data/demo_N/obs/language_instruction          (T,)  variable-length UTF-8 string
/data/demo_N/next_obs/language_instruction     (T,)  variable-length UTF-8 string

The instruction is constant within (and across) all demos of a task — see the table above.

---

Observation Space

Common keys (present in every task; image keys are uint8 RGB, shape (T, 256, 256, 3)):

Key	Shape	Description
obs/agentview_image	(T, 256, 256, 3)	Third-person camera
obs/robot0_eye_in_hand_image	(T, 256, 256, 3)	Wrist camera
obs/robot0_eef_pos	(T, 3)	End-effector position
obs/robot0_eef_quat	(T, 4)	End-effector quaternion
obs/robot0_gripper_qpos	(T, 2)	Gripper joint positions
obs/robot0_gripper_qvel	(T, 2)	Gripper joint velocities
obs/robot0_joint_pos	(T, 7)	Joint positions
obs/robot0_joint_vel	(T, 7)	Joint velocities
obs/language_instruction	(T,)	Task instruction (string)

PegInsertion additionally provides obs/peg_focus_view_image (T, 256, 256, 3). Some tasks also include robot0_eef_quat_site, robot0_joint_pos_cos, robot0_joint_pos_sin.

next_obs mirrors the non-image observation keys at timestep t+1. Image arrays are not stored under next_obs (to stay within the HF 50 GB per-file limit); reconstruct them at training time from obs/*_image shifted by one step if needed.

---

Action Space

Key	Dim	Description
actions	7	Concatenated: Δpos (3) + Δrot axis-angle (3) + gripper (1)
action_dict/rel_pos	3	Relative EEF translation
action_dict/rel_rot_axis_angle	3	Relative EEF rotation (axis-angle)
action_dict/rel_rot_6d	6	Relative EEF rotation (6D representation)
action_dict/gripper	1	Gripper open/close command

---

Data Format

HDF5 (Robosuite / RoboMimic convention). Each file contains a top-level /data group:

/data
  attrs: env_args (JSON), total (int)
  /demo_N
    attrs: num_samples, model_file, camera_info
    /obs/                  ← observations at t (incl. language_instruction)
    /next_obs/             ← observations at t+1 (non-image keys + language_instruction)
    /actions               ← (T, 7)
    /action_dict/          ← decomposed action components
    /rewards               ← (T,)
    /dones                 ← (T,)
    /states                ← Mujoco simulator states for replay
    /language_instruction  ← (T,) task instruction string
/mask                      ← train / valid demo splits (when present)

---

File Index

data/hdf5/
├── FruitSwapEasy/image_fruitswap_easy.hdf5
├── FruitSwapHard/image_fruitswap_hard.hdf5
├── PegInsertion/image_peginsertion.hdf5
└── UncoverBlock/image_uncoverblock.hdf5

---

Citation

@misc{memorymanipbench2026,
  title  = {Memory-Manip-Bench: 3D Contact-Rich Memory-Based Manipulation Tasks},
  year   = {2026},
  url    = {https://huggingface.co/datasets/harrywang01/Memory-Manip-Bench}
}