HDF5
Was this helpful?
Was this helpful?
Extend Robotics supports downloading datasets in HDF5 format — a versatile and efficient data structure widely used in machine learning, robotics, and scientific computing. You can learn more about the HDF5 format from the official .
This guide outlines the structure of the HDF5 dataset you receive from our platform.
The downloaded HDF5 dataset is organized into several key groups. Each group contains structured arrays or time-series data aligned by timestep, allowing easy parsing for downstream applications like training models or replaying episodes.
/observations
This group contains data collected from the robot's sensors and states at each timestep.
qpos: Joint positions, typically for the robot's gripper and arm.
qvel: Joint velocities, typically for the robot's gripper and arm..
effort: Efforts (torques/forces), typically for the robot's gripper and arm.
qpos_torso: Joint positions for the robot's torso.
qvel_torso: Joint velocities for the torso.
effort_torso: Efforts (torques/forces) specific to the torso's joints.
torso_action: Actions applied to the torso.
base_action: Velocity commands for the robot’s base.
contact_force: Gripper contact force feedback values.
images: Camera RGB images.
images_depth: Camera depth images.
/action
This group stores the actions executed by the robot at each timestep. Each action is represented as a concatenated array of joint commands for the arm, gripper, and (if applicable) torso or base components.