Cordelia/WheelArm_WoZ_Pilot_Dataset

WheelArm Synchronized Dataset

A multimodal dataset of wheelchair-mounted robot arm demonstrations for assistive daily-living tasks. Each episode captures a single task performed by a human operator and includes synchronized RGB video, depth, robot kinematics, audio, and natural-language dialogue with ambiguity annotations.

Dataset Summary

WheelArm is a real-robot dataset collected from a Kinova Gen3 6-DOF manipulator arm mounted on a powered wheelchair. Five subjects performed five assistive daily-living tasks across 53 episodes. Every episode provides temporally aligned streams from two RGB cameras, two depth cameras, all robot joint and Cartesian states, IMU, wheelchair base states, joystick commands, dual-microphone audio, and human-robot dialogue transcripts annotated for pragmatic ambiguity.

Property	Value
Total episodes	53
Task categories	5
Human subjects	5
Approx. total size	~47 GB
Audio sample rate	48 kHz mono PCM_16

---

Supported Tasks

Task	Episodes
drinking	9
door_opening	15
drawer_opening	16
cleaning	4
feeding	9

---

Dataset Structure

Directory layout

WheelArm_WoZ_Multimodal_Pilot/
├── drinking/
│   ├── 1-drinking-3/          # {subject}-{task}-{variant}
│   │   ├── cam_0_rgb_video.avi
│   │   ├── cam_0_rgb_video.metadata
│   │   ├── cam_0_depth.h5
│   │   ├── cam_2_rgb_video.avi
│   │   ├── cam_2_rgb_video.metadata
│   │   ├── cam_2_depth.h5
│   │   ├── cam_2_depth.metadata
│   │   ├── kinova_gen3_joint_states.h5
│   │   ├── kinova_gen3_cartesian_states.h5
│   │   ├── kinova_gen3_imu.h5
│   │   ├── kinova_gen3_wheelchair_states.h5
│   │   ├── kinova_gen3_wheelchair_joy_commands.h5
│   │   ├── headset_audio.wav
│   │   ├── headset_audio.metadata
│   │   ├── laptop_mic.wav
|   |   ├── laptop_mic.metadata
│   │   ├── synchronization/
|   |        ├──cam_0_synced_ref_fps.mp4
|   |        ├──cam_2_synced_ref_fps.mp4
|   |        ├──ee_jerk_stats.csv
|   |        ├──ee_jerk_timeseries.csv
|   |        ├──filtered_ee.csv
|   |        ├──filtered_joints.csv
|   |        ├──master.jsonl
|   |        ├──refgrid_interpolated_and_filtered.csv
|   |        ├──timestamps_synced_refgrid.csv
|   |     
│   ├── ...
│   └── summary/
├── door_opening/
├── drawer_opening/
├── cleaning/
└── feeding/

Episode naming

Episodes are named {subject}-{task}-{variant}:

• subject — integer 1–5, identifies the human operator
• task — abbreviated task name (drinking, door, drawer, cleaning, feeding)
• variant — integer repetition index within that subject × task pair

Example: 2-drinking-3 = Subject 2, drinking task, 3rd repetition.

---

Data Fields

RGB video (cam_0_rgb_video.avi, cam_2_rgb_video.avi)

Field	Value
Format	AVI
Cameras	cam_0 — ego view; cam_2 — wrist view
Frame rate	~12 Hz; ~15Hz

Accompanying .metadata files are Python pickle objects containing:

{
  "file_name": str,          # relative path to .avi
  "num_datapoints": int,     # total frames
  "record_start_time": float,  # Unix timestamp
  "record_end_time": float,
  "record_duration": float,  # seconds
  "record_frequency": float, # Hz
  "timestamps": list[float], # per-frame Unix timestamps
}

Depth data (cam_0_depth.h5, cam_2_depth.h5)

HDF5 files with per-frame depth arrays. Accompanying cam_2_depth.metadata also includes camera intrinsics:

{
  "num_datapoints": int,     # frames
  "record_frequency": float, # ~14.7–14.8 Hz
  "camera_info": {
    "resolution": [480, 270],  # width × height (pixels)
    "K": [...],              
    "distortion_model": "plumb_bob",
    "D": [0.0, 0.0, 0.0, 0.0, 0.0]
  }
}

Robot kinematics (kinova_gen3_*.h5)

All kinematic streams are HDF5 files with time-indexed arrays:

File	Contents
kinova_gen3_joint_states.h5	6 joint positions (rad), velocities (rad/s), efforts (N·m), timestamp (s)
kinova_gen3_cartesian_states.h5	End-effector position (m) + quaternion orientation
kinova_gen3_imu.h5	orientation (quaternion), orientation covariance, angular_velocity (rad/s), angular_velocity_covariance, linear_acceleration (m/s²), timestamp (s)
kinova_gen3_wheelchair_states.h5	left/right wheel angles (rad) and speeds
kinova_gen3_wheelchair_joy_commands.h5	axes, buttons, and timestamp (s)

Audio (headset_audio.wav, laptop_mic.wav)

Field	Value
Sample rate	48 000 Hz
Channels	1 (mono)
Bit depth	PCM_16
Codec frame	20 ms
Typical size	11–20 MB per file

Two microphones are provided per episode: a wearable headset microphone worn by the operator and a laptop microphone capturing the ambient scene.

Dialogue annotations (synchronization/ subdirectory)

Episodes with human-robot interaction include a synchronization/ folder containing:

File	Description
master.jsonl	Per-turn dialogue in conversational format with image references and ambiguity labels
frame_*.jpg	Key frames extracted for annotation (~18 per episode)
ee_jerk_stats.csv	End-effector jerk metrics (path length, mean/max jerk, jerk energy)
ee_jerk_timeseries.csv	End-effector jerk along x/y/z-axis, its magnitude and square
filtered_joints.csv	Filtered joint trajectories
filtered_ee.csv	Filtered end-effector trajectories
timestamps_synced_refgrid.csv	Reference-grid synchronisation timestamps
refgrid_interpolated_and_filtered.csv	data filtered with zero-phase 4th-order Butterworth
cam_0/2_synced_ref_fps	Synchronized videos following reference-grid timestamps