shuteng0608/forceawareact-peg-hole-mujoco

ForceAwareACT MuJoCo Peg-in-Hole Teleoperation Dataset

Dataset Summary

This dataset is a raw force-aware MuJoCo peg-in-hole teleoperation archive with synchronized robot state, RGB images, force/torque wrench streams, command-action labels, event records, and episode metadata. It contains 100 episode directories and 200 raw episode files. The raw episode archive inspected for this card is 13,344,579,333 bytes (12.43 GiB), excluding this card and generated audit reports.

Each episode directory is named with the pattern YYYYMMDD_HHMMSS_teleop_NNN and contains:

<episode_directory>/
  episode.hdf5
  metadata.json

All 100 inspected episodes share one HDF5 schema, and every episode contains both required files.

Dataset Description

The dataset was recorded from a MuJoCo wall peg-in-hole teleoperation setup. The HDF5 schema version stored in every file is compact_mujoco_hdf5_v1. The archive stores raw observations and metadata directly in HDF5. It is not a pre-converted Hugging Face datasets table, and the standard Dataset Viewer should not be assumed to render the HDF5 image tensors automatically.

The current ForceAwareACT training loader can consume this archive directly with h5py, derive fixed-length action/force chunks, align image and force streams to state timestamps, and compute normalization statistics.

Repository Structure

forceawareact-peg-hole-mujoco/
├── README.md
└── raw/
    ├── 20260701_131034_teleop_006/
    │   ├── episode.hdf5
    │   └── metadata.json
    └── ...

The Hugging Face repository stores the Dataset Card as the root README.md and the immutable episode archive under raw/. Local audit reports were used to prepare this card but are not required for loading the dataset.

Task Description

The task is a MuJoCo wall peg-in-hole task using a 7-DoF right-arm model with a cylindrical peg tool. The robot is teleoperated toward a wall fixture with a round hole. Episodes in this archive all have final HDF5 status auto_stop_task_success, and their events include task_success_site_reached, terminal_hold_start, and auto_stop_task_success.

Important caveat: the HDF5 attribute task_success is unknown in all inspected files. For this dataset, success-stop status is represented by the file/group status attributes and event records, not by the task_success attribute.

Data Collection

The matching recorder implementation is a compact MuJoCo HDF5 recorder. It records:

• state stream at nominal 30 Hz;
• image stream at nominal 30 Hz;
• force/torque stream at nominal 500 Hz;
• command-action labels aligned with state samples;
• episode events;
• initial and final robot/task metadata.

The recorder uses MuJoCo data.time as the alignment clock. The sidecar metadata.json stores only summary fields; detailed arrays and most metadata live in episode.hdf5.

The collection configuration inspected alongside this archive disables hole randomization (enable_hole_randomization: false and randomize_hole_on_record_start: false). No hole-randomization fields are present in metadata.json.

Sensor Modalities

Modality	Stored paths	Rate summary
Joint state	observations/joint_pos, observations/joint_vel, observations/joint_torque	approximately 30 Hz
End-effector pose	observations/ee_pose	approximately 30 Hz
RGB cameras	observations/images/ee_cam, observations/images/base_top_cam	approximately 30 Hz
Force/torque wrench	observations/ft_wrench, observations/ft_wrench_raw, observations/ft_wrench_gravity	approximately 500 Hz
Command actions	actions/joint_pos_command, action	approximately 30 Hz
Events	events/names, events/t_episode, events/t_sim, events/t_wall	event records, not periodic samples

HDF5 Schema

Every inspected episode.hdf5 file contains these dataset paths:

• action
• actions/joint_pos_command
• episode_metadata/actuator_names
• episode_metadata/camera_names
• episode_metadata/final_ee_pose
• episode_metadata/final_ft_wrench
• episode_metadata/final_ft_wrench_gravity
• episode_metadata/final_ft_wrench_raw
• episode_metadata/final_hole_center_pos
• episode_metadata/final_joint_pos
• episode_metadata/final_joint_pos_command
• episode_metadata/final_joint_torque
• episode_metadata/final_joint_vel
• episode_metadata/final_peg_tip_pos
• episode_metadata/ft_gravity_tool_body_names
• episode_metadata/ft_gravity_tool_com_sensor_initial
• episode_metadata/ft_gravity_tool_com_world_initial
• episode_metadata/ft_gravity_world
• episode_metadata/initial_ee_pose
• episode_metadata/initial_ft_wrench
• episode_metadata/initial_ft_wrench_gravity
• episode_metadata/initial_ft_wrench_raw
• episode_metadata/initial_hole_center_pos
• episode_metadata/initial_joint_pos
• episode_metadata/initial_joint_pos_command
• episode_metadata/initial_joint_torque
• episode_metadata/initial_joint_vel
• episode_metadata/initial_peg_tip_pos
• episode_metadata/initial_task_error_xyz
• episode_metadata/joint_names
• events/names
• events/t_episode
• events/t_sim
• events/t_wall
• observations/ee_pose
• observations/ft_wrench
• observations/ft_wrench_gravity
• observations/ft_wrench_raw
• observations/images/base_top_cam
• observations/images/camera_names
• observations/images/ee_cam
• observations/joint_pos
• observations/joint_torque
• observations/joint_vel
• timestamps/force
• timestamps/force_episode
• timestamps/image
• timestamps/image_episode
• timestamps/state
• timestamps/state_episode

Field Summary

HDF5 path	Episodes	Shape summary	dtype	Compression	Chunks
action	100/100	varies by episode; first-dim min 213, max 448, mean 309.77, median 306.50	float64	none	(256, 7)
actions/joint_pos_command	100/100	varies by episode; first-dim min 213, max 448, mean 309.77, median 306.50	float64	none	(256, 7)
episode_metadata/actuator_names	100/100	(7,) x100	object	none	none
episode_metadata/camera_names	100/100	(2,) x100	object	none	none
episode_metadata/final_ee_pose	100/100	(7,) x100	float64	none	none
episode_metadata/final_ft_wrench	100/100	(6,) x100	float64	none	none
episode_metadata/final_ft_wrench_gravity	100/100	(6,) x100	float64	none	none
episode_metadata/final_ft_wrench_raw	100/100	(6,) x100	float64	none	none
episode_metadata/final_hole_center_pos	100/100	(3,) x100	float64	none	none
episode_metadata/final_joint_pos	100/100	(7,) x100	float64	none	none
episode_metadata/final_joint_pos_command	100/100	(7,) x100	float64	none	none
episode_metadata/final_joint_torque	100/100	(7,) x100	float64	none	none
episode_metadata/final_joint_vel	100/100	(7,) x100	float64	none	none
episode_metadata/final_peg_tip_pos	100/100	(3,) x100	float64	none	none
episode_metadata/ft_gravity_tool_body_names	100/100	(1,) x100	object	none	none
episode_metadata/ft_gravity_tool_com_sensor_initial	100/100	(3,) x100	float64	none	none
episode_metadata/ft_gravity_tool_com_world_initial	100/100	(3,) x100	float64	none	none
episode_metadata/ft_gravity_world	100/100	(3,) x100	float64	none	none
episode_metadata/initial_ee_pose	100/100	(7,) x100	float64	none	none
episode_metadata/initial_ft_wrench	100/100	(6,) x100	float64	none	none
episode_metadata/initial_ft_wrench_gravity	100/100	(6,) x100	float64	none	none
episode_metadata/initial_ft_wrench_raw	100/100	(6,) x100	float64	none	none
episode_metadata/initial_hole_center_pos	100/100	(3,) x100	float64	none	none
episode_metadata/initial_joint_pos	100/100	(7,) x100	float64	none	none
episode_metadata/initial_joint_pos_command	100/100	(7,) x100	float64	none	none
episode_metadata/initial_joint_torque	100/100	(7,) x100	float64	none	none
episode_metadata/initial_joint_vel	100/100	(7,) x100	float64	none	none
episode_metadata/initial_peg_tip_pos	100/100	(3,) x100	float64	none	none
episode_metadata/initial_task_error_xyz	100/100	(3,) x100	float64	none	none
episode_metadata/joint_names	100/100	(7,) x100	object	none	none
events/names	100/100	(5,) x100	object	none	none
events/t_episode	100/100	(5,) x100	float64	none	none
events/t_sim	100/100	(5,) x100	float64	none	none
events/t_wall	100/100	(5,) x100	float64	none	none
observations/ee_pose	100/100	varies by episode; first-dim min 213, max 448, mean 309.77, median 306.50	float64	none	(256, 7)
observations/ft_wrench	100/100	varies by episode; first-dim min 3548, max 7455, mean 5155.24, median 5106.00	float64	none	(2048, 6)
observations/ft_wrench_gravity	100/100	varies by episode; first-dim min 3548, max 7455, mean 5155.24, median 5106.00	float64	none	(2048, 6)
observations/ft_wrench_raw	100/100	varies by episode; first-dim min 3548, max 7455, mean 5155.24, median 5106.00	float64	none	(2048, 6)
observations/images/base_top_cam	100/100	varies by episode; first-dim min 213, max 448, mean 309.77, median 306.50	uint8	lzf	(1, 480, 640, 3)
observations/images/camera_names	100/100	(2,) x100	object	none	none
observations/images/ee_cam	100/100	varies by episode; first-dim min 213, max 448, mean 309.77, median 306.50	uint8	lzf	(1, 480, 640, 3)
observations/joint_pos	100/100	varies by episode; first-dim min 213, max 448, mean 309.77, median 306.50	float64	none	(256, 7)
observations/joint_torque	100/100	varies by episode; first-dim min 213, max 448, mean 309.77, median 306.50	float64	none	(256, 7)
observations/joint_vel	100/100	varies by episode; first-dim min 213, max 448, mean 309.77, median 306.50	float64	none	(256, 7)
timestamps/force	100/100	varies by episode; first-dim min 3548, max 7455, mean 5155.24, median 5106.00	float64	none	(2048,)
timestamps/force_episode	100/100	varies by episode; first-dim min 3548, max 7455, mean 5155.24, median 5106.00	float64	none	(2048,)
timestamps/image	100/100	varies by episode; first-dim min 213, max 448, mean 309.77, median 306.50	float64	none	(1,)
timestamps/image_episode	100/100	varies by episode; first-dim min 213, max 448, mean 309.77, median 306.50	float64	none	(1,)
timestamps/state	100/100	varies by episode; first-dim min 213, max 448, mean 309.77, median 306.50	float64	none	(256,)
timestamps/state_episode	100/100	varies by episode; first-dim min 213, max 448, mean 309.77, median 306.50	float64	none	(256,)

All numeric state/action/force datasets are float64. Camera tensors are uint8 RGB, LZF-compressed, and chunked as one full frame per chunk.

metadata.json Schema

Every sidecar metadata.json has these keys:

Key	Episodes	Type	Notes
camera_names	100/100	list	['ee_cam', 'base_top_cam']
duration_sim	100/100	float	simulation duration in seconds
episode_label	100/100	string	e.g. teleop_006
hdf5_path	100/100	string	original recorder-machine path; not portable
image_storage	100/100	string	inside_hdf5_uint8_rgb
n_force	100/100	integer	force sample count
n_image	100/100	integer	image sample count
n_state	100/100	integer	state sample count
schema_version	100/100	string	compact_mujoco_hdf5_v1
status	100/100	string	auto_stop_task_success

metadata.json does not contain task geometry, hole-randomization fields, full timestamp arrays, force streams, images, or initial/final robot arrays. Those are stored in episode.hdf5.

Temporal Sampling

Measured from the timestamp arrays across all 100 episodes:

Stream	Samples per episode	Duration seconds	Measured rate Hz	Timestamp check
State, timestamps/state_episode	min 213, max 448, mean 309.77, median 306.50	min 7.066, max 14.900, mean 10.292, median 10.183	min 29.996, max 30.004, mean 30.001, median 30.001	finite and strictly increasing
Image, timestamps/image_episode	min 213, max 448, mean 309.77, median 306.50	min 7.066, max 14.900, mean 10.292, median 10.183	min 29.996, max 30.004, mean 30.001, median 30.001	finite and strictly increasing
Force, timestamps/force_episode	min 3548, max 7455, mean 5155.24, median 5106.00	min 7.093, max 14.908, mean 10.308, median 10.209	min 499.938, max 500.070, mean 500.027, median 500.043	finite and strictly increasing
Events, events/t_episode	5 records in every episode	min 7.094, max 14.910, mean 10.309, median 10.210	not periodic	finite and nondecreasing; duplicate event times occur

State, image, and force streams begin at the same episode time in every episode. Their end times differ by at most 0.032 seconds across streams.

Action Definition

The raw HDF5 archive contains explicit command-action labels in every episode:

• actions/joint_pos_command: [N_state, 7], actual MuJoCo position-actuator command from data.ctrl[actuator_ids].
• action: [N_state, 7], ACT-compatible alias of actions/joint_pos_command.

The ForceAwareACT loader also supports training with action_mode="joint_pos", where labels are derived from future observed joint positions:

action_chunk = observations/joint_pos[i + 1 : i + K + 1]

For command modes such as action and joint_pos_command, chunks are aligned to the current decision index:

command_chunk = command[i : i + K]

Delta action modes subtract the current joint position from the selected command chunk. The raw dataset does not define a fixed chunk length; K is a loader/training argument. Samples near the end of an episode that cannot provide a full future chunk are excluded rather than padded.

Force/Torque Representation

The wrench component order is:

[Fx, Fy, Fz, Tx, Ty, Tz]

Every episode stores:

• observations/ft_wrench_raw: raw MuJoCo force/torque sensor reading;
• observations/ft_wrench_gravity: predicted gravity wrench in the sensor frame;
• observations/ft_wrench: compensated wrench used by default for learning.

For this dataset, HDF5 metadata records:

observations/ft_wrench = observations/ft_wrench_raw - observations/ft_wrench_gravity

The compensation mode is gravity; episode-start tare is disabled. Recorder code expresses the gravity wrench in the FT sensor site frame using the configured tool body and gravity vector. A formal public frame diagram and sign-convention statement are still required before public release.

During ForceAwareACT training, causal force windows are derived from observations/ft_wrench by sampling only force timestamps less than or equal to the current state timestamp. Future force chunks for force-prediction losses are aligned to future state timestamps by nearest force timestamp.

Camera Streams

The archive contains two camera streams in every episode:

Camera	HDF5 path	Shape	dtype	Compression	Verified role
ee_cam	observations/images/ee_cam	[N_image, 480, 640, 3]	uint8 RGB	LZF	end-effector camera attached near the flange/peg connection
base_top_cam	observations/images/base_top_cam	[N_image, 480, 640, 3]	uint8 RGB	LZF	fixed top/base camera looking at the wall task region

Sampled HDF5 frames show rendered MuJoCo RGB images without text HUD overlays. The force-feedback HUD is drawn only on live OpenCV display frames by the controller code and is not recorded into these HDF5 image tensors.

Camera calibration matrices and intrinsics are not stored in the HDF5 files.

Episode Metadata

HDF5 episode_metadata contains:

• names of joints, actuators, cameras, and force-gravity tool bodies;
• initial/final joint position, velocity, torque, and end-effector pose;
• initial/final force wrench, raw wrench, and gravity wrench;
• initial/final peg-tip positions;
• force compensation attributes;
• sample counts and recording status.

Legacy hole-goal metadata caveat

The active task-success target in the current MuJoCo model is hole_goal_site. However, the recorder metadata embedded in this historical 100-episode archive refers to the legacy site name hole_center_site. That legacy name did not resolve when the archived metadata was written.

Consequently, episode_metadata/initial_hole_center_pos, episode_metadata/final_hole_center_pos, and episode_metadata/initial_task_error_xyz contain non-finite values in all 100 episodes. These fields must not be used as valid hole-position or task-error labels. Recorded task completion should instead be interpreted from the status attributes and event records, including task_success_site_reached and auto_stop_task_success.

Loading Example

Use h5py and load only the slices you need, especially for image tensors.

from pathlib import Path
import h5py

root = Path("raw")
episode = sorted(root.glob("*/episode.hdf5"))[0]

with h5py.File(episode, "r") as f:
    qpos = f["observations/joint_pos"][:]          # small numeric array
    force_window_source = f["observations/ft_wrench"]
    first_force_sample = force_window_source[0]

    # Avoid loading the full image tensor unless you really need it.
    ee_first_frame = f["observations/images/ee_cam"][0]

    print(f.attrs["schema_version"])
    print(qpos.shape, first_force_sample.shape, ee_first_frame.shape)

For ForceAwareACT training, prefer using the repository loader rather than hand-aligning streams.

Download Instructions

Install the Hugging Face CLI, authenticate if the dataset is private, and download the raw archive:

python -m pip install -U huggingface_hub
hf auth login

hf download \
  shuteng0608/forceawareact-peg-hole-mujoco \
  --repo-type dataset \
  --local-dir forceawareact-peg-hole-mujoco

After download, run the loading example from the repository root, where the episode archive is located under raw/.

Intended Uses

• Research on imitation learning for contact-rich manipulation in simulation.
• Force-aware policy learning with synchronized vision, proprioception, and force/torque observations.
• Offline evaluation of action prediction and future force prediction models.
• Reproducibility studies for the ForceAwareACT data pipeline.

Out-of-Scope Uses

• Direct real-world robot deployment or safety validation.
• Claims about real sensor noise, real calibration, or physical hardware safety.
• Training methods that require frame-level human intent/correction annotations unless those annotations are added separately.
• Treating this raw HDF5 archive as a Dataset Viewer-native image dataset without conversion.

Known Limitations

• The dataset is simulation-only MuJoCo data.
• Public license/access terms are not defined in the repository at audit time.
• Camera intrinsics/calibration are not stored in HDF5.
• The formal force/torque frame and sign convention need a public diagram or specification.
• metadata.json contains original recorder-machine hdf5_path strings and should not be used as portable paths.
• The task_success attribute on episode_metadata is unknown; success-stop information is in status and event records.
• The historical recorder metadata refers to the legacy site name hole_center_site, while the active task-success target is hole_goal_site. Consequently, the archived hole-center position and task-error fields are non-finite and must not be used as valid labels.
• Human corrective behavior during contact is not explicitly annotated.

Dataset Versioning

This card describes the inspected raw archive with HDF5 schema version compact_mujoco_hdf5_v1 and 100 episode directories. Future releases should update this card when episodes are added, schema changes, or license/access terms are clarified.

Integrity Verification

Expected raw archive properties for this version:

• Episode directories: 100
• Raw episode files: 200
• Raw archive size: 13,344,579,333 bytes (12.43 GiB)
• Per-episode required files: episode.hdf5, metadata.json
• HDF5 schema variants: 1
• HDF5 read errors in audit: 0
• Metadata JSON read errors in audit: 0

A minimal verification command after download:

python - <<'PYCODE'
from pathlib import Path
root = Path('raw')
episodes = sorted(p for p in root.iterdir() if p.is_dir())
print('episodes', len(episodes))
print('missing required files', [p.name for p in episodes if not (p/'episode.hdf5').is_file() or not (p/'metadata.json').is_file()])
PYCODE

License and Access

No repository license file was found during this audit. Do not assume a public license until the dataset owner publishes one. Access restrictions, redistribution terms, and citation requirements should be defined before public release.

Citation

No paper citation is defined by the repository at audit time. If you use this dataset, cite the dataset repository and any future associated paper or project page provided by the authors.

Documentation Still Required Before Public Release

• Public license and access policy.
• Formal wrench coordinate-frame and sign-convention documentation.
• Camera calibration/intrinsics policy, or an explicit statement that calibration is unavailable.
• Definition of a stable public version tag or release revision.
• Clarification of whether public consumers should use raw HDF5 directly or a converted datasets package.