Hugging Face's logo

Datasets:
nvidia
/
PhysicalAI-Robotics-Open-H-Embodiment

Tasks:
Robotics
Modalities:
Tabular
Text
Video
Size:
< 1K
Tags:
robotics
healthcare
License:
Dataset card Data Studio Files
xet
 Community
2
PhysicalAI-Robotics-Open-H-Embodiment / Surgical /hamlyn /peg_transfer /README.md
nigeln's picture
nigeln
Add mapped dataset README files
40e876a verified
preview code
|
Raw
History Blame Contribute Delete
10.2 kB

Hamlyn-dVRK: Table-Top Peg Transfer - README

At a Glance

High-fidelity teleoperated demonstrations of a bimanual da Vinci Research Kit (dVRK) robot performing peg transfer on a table-top training rig. The robot uses DeBakey forceps (left) and a needle driver (right) to pick up pegs, perform a handover between tools, and place pegs onto specified posts.

Dataset Overview

This dataset contains teleoperated trajectories of trained operators using the dVRK to perform peg transfer on a table-top training apparatus. The goal is to pick up a peg with the left hand, transfer it to the right hand, and place it accurately onto a target post according to a given instruction.

Task Logic & Execution: The operator utilises the Patient Side Manipulators (PSMs) equipped with DeBakey forceps (left) and a needle driver (right).

  • Pick (Left Hand): The left hand grasps the instructed peg.
  • Handover + Place (Right Hand): The peg is transferred to the right hand and placed onto a target post on the board.

Key Features:

  • Table-Top Dexterity Primitive: Clean visual scene with rigid objects, supporting accurate benchmarking of bimanual handovers and placement.
  • Instruction-Conditioned Placement: Demonstrations follow a given instruction specifying which peg to move and where to place it.
Total Trajectories 317
Total Hours 0.545
Data Type [ ] Clinical [ ] Ex-Vivo [x] Table-Top Phantom [ ] Digital Simulation [ ] Physical Simulation [ ] Other
License CC BY 4.0
Version 1.0

Hamlyn dVRK Dataset Overview

This is part of the Hamlyn dVRK Dataset, which encompasses six distinct surgical tasks. All data were collected through teleoperation on bimanual tasks performed on ex-vivo animal samples and phantoms.

Task Conditioning

The tasks are categorized into two types:

  • Non-conditioned: Single sentence task descriptions
  • Language-conditioned: Multiple variants of task descriptions across episodes

Dataset Statistics

  • Total Episodes: 972
  • Total Frames: 545k @ 30Hz
  • Dataset Duration: 5.04 hours

Episode Outcomes

Outcome Category Episodes Description
Success 780 Task completed successfully with specified conditions
Recovery 110 Task finished with recovery behavior
Failed 82 Task failed to complete

*Success rates vary across tasks due to varying difficulty levels for operators during teleoperation.

Tasks & Domain

Domain

Select the primary domain for this dataset.

  • Surgical Robotics
  • Ultrasound Robotics
  • Other Healthcare Robotics

Demonstrated Skills

This specific dataset subset focuses on Peg Transfer.

Note: This is part of the Hamlyn dVRK Data Collection which also includes:

  • Tissue Retraction / Exposure
  • Knot Tying
  • Suturing (Single Loop)
  • Suturing (Dual Loop)
  • Peg Transfer
  • Needle Grasp and Handover

Data Collection Details

Collection Method

How was the data collected?

  • Human Teleoperation
  • Programmatic/State-Machine
  • AI Policy / Autonomous
  • Other

Operator Details

Description
Operator Count 2
Operator Skill Level [ ] Expert (e.g., Surgeon, Sonographer)
[x] Intermediate (e.g., Trained Researcher)
[x] Novice (e.g., ML Researcher with minimal experience)
[ ] N/A
Collection Period From 2025-12-01 to 2026-01-15
Input Interface Teleoperation interface with bimanual Force Dimension Sigma 7 haptic device

Recovery Demonstrations

Does this dataset include examples of recovering from failure?

  • Yes
  • No

If yes, please briefly describe the recovery process:

In the peg transfer task, failure modes typically involve dropping the peg during transport, an unstable handover, or misplacement onto the post. In these instances, the operator does not abort the episode but performs a recovery manoeuvre: re-grasping the peg, re-aligning the approach, and repeating the handover and placement steps.

Diversity Dimensions

Check all dimensions that were intentionally varied during data collection.

  • Camera Position / Angle
  • Lighting Conditions
  • Target Object (Varied peg and target post)
  • Spatial Layout (Varied board pose / peg arrangement)
  • Robot Embodiment
  • Task Execution (Left vs. Right hand dominance / Approach angle)
  • Background / Scene
  • Initial Robot Configuration (Table-top domain randomisation)

Elaboration on Diversity:

  • Target Object: A table-top peg board was used. Diversity is achieved through varying which peg is selected and which post is the placement target.
  • Spatial Layout: The board pose and/or peg arrangement are varied between sets to reduce overfitting to absolute coordinates.
  • Task Execution: Operators vary approach angle, grasp point, and handover strategy to maintain stable transport and accurate placement.
  • Initial Robot Configuration: End-effectors were reset to a randomised "home" position above the workspace before each trajectory.

🛠️ Equipment & Setup

Robotic Platform(s)

List the primary robot(s) used.

  • Robot 1: dVRK (da Vinci Research Kit) - Patient Side Manipulators (PSM1, PSM2)

Sensors & Cameras

List the sensors and cameras used. Specify model names where possible.

Type Model/Details
Primary Camera Intel RealSense D405 (RGB + Depth), 848x480 @ 30fps
Wrist Camera INSKAM 5.5mm USB Endo-Cam (Right/Left Arm) with custom built mount, view angle 66 degree, focal length 4-10cm, 640x480 @ 24fps
Kinematics dVRK High-Frequency Joint Encoders (100Hz)

🎯 Action & State Space Representation

The dataset follows the standard LeRobot format for bimanual manipulation.

Action Space Representation

Primary Action Representation:

  • Absolute Cartesian (position/orientation relative to robot base)
  • Relative Cartesian
  • Joint Space

Orientation Representation:

  • Quaternions (x, y, z, w)
  • Euler Angles
  • Rotation Matrix

Reference Frame:

  • Robot Base Frame (Base of each PSM arm)
  • Camera Frame

Action Dimensions:

action: [x, y, z, qx, qy, qz, qw, gripper_angle, x, y, z, qx, qy, qz, qw, gripper_angle]

  • The first 8 dimensions are for the left arm and last 8 dimensions are for the right arm
  • x, y, z: Absolute position in PSM base frame (meters)
  • qx, qy, qz, qw: Absolute orientation as quaternion
  • gripper_angle: Normalized jaw opening angle (1 = opened, 0 = closed)

State Space Representation

State Information Included:

  • Joint Positions
  • Joint Velocities
  • End-Effector Pose (No extra end-effector used with dVRK)
  • Gripper State

Primary State:

observation.state: [x, y, z, qx, qy, qz, qw, gripper_angle, x, y, z, qx, qy, qz, qw, gripper_angle]

  • The first 8 dimensions are for the left arm and last 8 dimensions are for the right arm
  • x, y, z: Absolute position in PSM base frame (meters)
  • qx, qy, qz, qw: Absolute orientation as quaternion
  • gripper_angle: Normalized jaw opening angle (1 = opened, 0 = closed)

State Dimensions:

observation.state.left_arm_cartesian: [x, y, z, qx, qy, qz, qw, gripper_angle]

observation.state.right_arm_cartesian: [x, y, z, qx, qy, qz, qw, gripper_angle]

observation.state.left_arm_joint: [j1, j2, j3, j4, j5, j6, gripper_angle]

observation.state.right_arm_joint: [j1, j2, j3, j4, j5, j6, gripper_angle]

  • j1-j6: Joint positions for 6-DOF PSM arm (radians)
  • gripper_angle: Normalized jaw opening angle (1 = opened, 0 = closed)

⏱️ Data Synchronization Approach

Describe how you achieved proper data synchronization.

Distributed Synchronization Architecture: The data collection setup involves a distributed system with a robot control machine that teleop the dVRK and a sensor recording machine that stream kinematics, and videos from camers.

  1. Clock Sync: We utilize Chrony (NTP) to enforce strict time alignment between the two workstations, minimizing clock skew to < 3.3 ms.
  2. Timestamping: All data streams (camera RGB and depth images, wrist camer images, robot kinematics) are timestamped with ROS wall-time headers at the exact moment of capture.
  3. Delay Compensattion and Trimming: The different cameras can introduce small relative delays. These delays are first estimated and compensated by applying a common temporal offset to each stream. Subsequently, all streams are temporally trimmed to their common interval, bounded by the latest start time and the earliest end time across all modalities.
  4. Alignment: During the LeRobot conversion process, frames from all data streams are aligned to a reference timeline at 30 Hz30 Hz. This is achieved via nearest-neighbor temporal interpolation based on the synchronized ROS timestamps.

Attribution & Contact

Please provide attribution for the dataset creators and a point of contact.

Dataset Lead Jianhao Su, Kaizhong Deng, Zhaoyang Jacopo Hu
Institution The Hamlyn Centre for Robotic Surgery, Imperial College London
Contact Email k.deng21@imperial.ac.uk (for technical issues), stamatia.giannarou@imperial.ac.uk (for correspondence)
Citation (BibTeX) 
@misc{hamlyn_dvrk_openh_2026,
  author = {Su, Jianhao and Deng, Kaizhong and Hu, Zhaoyang Jacopo and Rodriguez y Baena, Ferdinando and Giannarou, Stamatia},
  title = {Hamlyn dVRK Surgical Manipulation Dataset for Open-H},
  year = {2026},
  publisher = {Open-H-Embodiment},
  note = {Table-Top Domain: Peg Transfer}
}