Dataset Card for VLA_pick_and_place

Dataset Summary

This dataset contains 70 teleoperated demonstration episodes across 4 pick-and-place style manipulation tasks, collected on the Garmi dual-arm robot (two Franka arms with Franka Hand grippers). It was recorded to train/evaluate Vision-Language-Action (VLA) manipulation policies.

Each episode is a raw ROS 2 mcap rosbag capturing synchronized RGB camera streams and full-arm proprioception (joint states, end-effector poses, twists, wrenches) for both arms.

Task	Episodes	Total duration	Notes
grasp_bottle	55 (garmi_demo_001–005, 021–070) (see note)	~32.3 min	Grasp a bottle
grasp_cup	5 (garmi_demo_006–010)	~4.8 min	Grasp a cup
grasp_objects_to_box	5 (garmi_demo_011–015)	~4.9 min	Grasp objects and place them into a box
grasp_objects_to_box_and_move_box	5 (garmi_demo_016–020)	~6.9 min	Grasp objects into a box, then move the box

Total: 70 episodes, ~49 minutes, ~61 GB. Recording dates: 2026-08-12, 2026-09-08, 2026-09-09. ROS distro: Jazzy.

Note: garmi_demo_001–005 and garmi_demo_021–070 are both grasp_bottle, recorded in separate sessions (2026-08-12 vs. 2026-09-08/09) — episode numbering is not contiguous because 006–020 belong to the other three tasks recorded in between.

Dataset Structure

VLA_pick_and_place/
├── grasp_bottle/
│   ├── garmi_demo_001/ ... garmi_demo_005/     # 2026-08-12 session
│   └── garmi_demo_021/ ... garmi_demo_070/     # 2026-09-08/09 sessions
├── grasp_cup/
│   └── garmi_demo_006/ ... garmi_demo_010/
├── grasp_objects_to_box/
│   └── garmi_demo_011/ ... garmi_demo_015/
└── grasp_objects_to_box_and_move_box/
    └── garmi_demo_016/ ... garmi_demo_020/

Each episode folder contains:

• garmi_demo_XXX_0.mcap — the ROS 2 bag (all topics below)
• metadata.yaml — standard rosbag2 bag metadata (topics, durations, message counts, QoS)

Recorded topics

Topic	Type	Approx. rate	Description
/camera_d435i_left/.../color/image_raw/compressed	sensor_msgs/CompressedImage	~30 Hz	Left wrist RealSense D435i RGB
/camera_d435i_right/.../color/image_raw/compressed	sensor_msgs/CompressedImage	~30 Hz	Right wrist RealSense D435i RGB
/camera_d435i_{left,right}/.../compressedDepth	sensor_msgs/CompressedImage	~30 Hz	Wrist depth — only present in garmi_demo_001, not the rest of the dataset
/olive/olixO1/id004/image/compressed	sensor_msgs/CompressedImage	~24 Hz	Scene/head camera
/garmi/{left,right}/franka_gripper/joint_states	sensor_msgs/JointState	~15 Hz	Gripper finger joint states
/garmi/{left,right}/franka_robot_state_broadcaster/measured_joint_states	sensor_msgs/JointState	~985 Hz	Measured arm joint state
/garmi/{left,right}/franka_robot_state_broadcaster/desired_joint_states	sensor_msgs/JointState	~985 Hz	Commanded arm joint state
/garmi/{left,right}/franka_robot_state_broadcaster/current_pose	geometry_msgs/PoseStamped	~985 Hz	Measured end-effector pose
/garmi/{left,right}/franka_robot_state_broadcaster/last_desired_pose	geometry_msgs/PoseStamped	~985 Hz	Commanded end-effector pose
/garmi/{left,right}/franka_robot_state_broadcaster/desired_end_effector_twist	geometry_msgs/TwistStamped	~985 Hz	Commanded end-effector twist
/garmi/{left,right}/franka_robot_state_broadcaster/external_wrench_in_base_frame	geometry_msgs/WrenchStamped	~985 Hz	External wrench, base frame
/garmi/{left,right}/franka_robot_state_broadcaster/external_wrench_in_stiffness_frame	geometry_msgs/WrenchStamped	~985 Hz	External wrench, stiffness frame
/garmi/{left,right}/franka_robot_state_broadcaster/external_joint_torques	sensor_msgs/JointState	~985 Hz	External joint torque estimate
/garmi/{left,right}/franka_robot_state_broadcaster/transition_event	lifecycle_msgs/TransitionEvent	event-driven	Controller lifecycle transitions

Both arms (left/right) publish the same set of proprioceptive topics.

Data Collection

Episodes were collected via human teleoperation: an operator remotely controlled the Garmi robot's arms to perform each task while ros2 bag record captured the topics above.

Each recording followed a stop-cue procedure: the operator started the bag, performed the task, and signaled a verbal "stop" cue; recording then continued for an additional 7 seconds after the cue before being terminated, so the tail end of each episode (e.g. task completion) is fully captured.

How to Use

The data is raw mcap/ROS 2 bags, not a pre-packaged datasets-library format. Download the files and read them with the mcap or rosbags Python packages:

pip install huggingface_hub mcap mcap-ros2-support

from huggingface_hub import snapshot_download
from mcap_ros2.reader import read_ros2_messages

local_dir = snapshot_download("danielkha/VLA_pick_and_place", repo_type="dataset")

for msg in read_ros2_messages(f"{local_dir}/grasp_bottle/garmi_demo_021/garmi_demo_021_0.mcap"):
    print(msg.channel.topic, msg.ros_msg)

License

Not yet finalized. Until a license is set, please contact the dataset owner before reusing this data outside of personal/research inspection.

Citation

If you use this dataset, please reach out to the owner (Hugging Face: danielkha) for citation guidance, as this is part of an in-progress Master's thesis project.