HORA: Hand–Object to Robot Action Dataset

Dataset Summary

HORA (Hand–Object to Robot Action) is a large-scale multimodal dataset that converts human hand–object interaction (HOI) demonstrations into robot-usable supervision for cross-embodiment learning. It combines HOI-style annotations (e.g., MANO hand parameters, object pose, contact) with embodied-robot learning signals (e.g., robot observations, end-effector trajectories) under a unified canonical action space.

HORA is constructed from three sources/subsets:

1. HORA(Mocap): custom multi-view motion capture system with tactile-sensor gloves (includes tactile maps).
2. HORA(Recordings): custom RGB(D) HOI recording setup (no tactile).
3. HORA(Public Dataset): derived from multiple public HOI datasets and retargeted to robot embodiments (6/7-DoF arms).

Overall scale: ~150k trajectories across all subsets.

---

Key Features

• Unified multimodal representation across subsets, covering both HOI analysis and downstream robotic learning.
• HOI modalities: MANO hand parameters (pose/shape + global transform), object 6DoF pose, object assets, hand–object contact annotations.
• Robot modalities: wrist-view & third-person observations, and end-effector pose trajectories for robotic arms, all mapped to a canonical action space.
• Tactile (mocap subset): dense tactile map for both hand and object (plus object pose & assets).

---

Dataset Statistics

Subset	Tactile	#Trajectories	Notes
HORA(Mocap)	✅	63,141	6-DoF object pose + assets + tactile map
HORA(Recordings)	❌	23,560	6-DoF object pose + assets
HORA(Public Dataset)	❌	66,924	retargeted cross-embodiment robot modalities
Total		~150k

Supported Tasks and Use Cases

HORA is suitable for:

• Imitation Learning (IL) / Visuomotor policy learning
• Vision–Language–Action (VLA) model training and evaluation
• HOI-centric research: contact analysis, pose/trajectory learning, hand/object dynamics

---

Data Format

Example Episode Structure

Each episode/trajectory may include:

HOI fields

• hand_mano: MANO parameters (pose/shape, global rotation/translation)
• object_pose_6d: 6DoF object pose sequence
• contact: hand–object contact annotations
• object_asset: mesh/texture id or path

Robot fields

• Global Attributes

  • task_description: Natural language instruction for the task (stored as HDF5 attribute).
  • total_demos: Total number of trajectories in the file.

• Observations (obs group)

  • agentview_rgb: JPEG byte stream (variable length uint8). Decodes to (T, 480, 640, 3).
  • eye_in_hand_{side}_rgb: JPEG byte stream (variable length uint8). Decodes to (T, 480, 640, 3).
  • {prefix}_joint_states: Arm joint positions in radians. Shape (T, N_dof).
  • {prefix}_gripper_states: Gripper joint positions. Shape (T, N_grip).
  • {prefix}_eef_pos: End-effector position in Robot Base Frame. Shape (T, 3).
  • {prefix}_eef_quat: End-effector orientation (w, x, y, z) in Robot Base Frame. Shape (T, 4).
  • object_{name}_pos: Object ground truth position in World Frame. Shape (T, 3).
  • object_{name}_quat: Object ground truth orientation (w, x, y, z) in World Frame. Shape (T, 4).

• Actions & States

  Note: For multi-robot setups, the fields below concatenate data from all robots in order (e.g., [robot0, robot1]).

  • actions: Joint-space control targets. Shape (T, N_dof + 1). Format: [joint_positions, normalized_gripper] where gripper is in [0, 1].
  • actions_ee: Cartesian control targets. Shape (T, 7). Format: [pos (3), axis-angle (3), normalized_gripper (1)].
  • robot_states: Robot base pose in World Frame. Shape (T, 7 * N_robots). Format: [pos (3), quat (4)] per robot, quat is (w, x, y, z).

Tactile fields (mocap only)

• tactile_hand: dense tactile map (time × sensors/vertices)
• tactile_object: dense tactile map

---