metadata
license: cc-by-nc-4.0
pretty_name: Lidar-Tactile — Raw ROS 2 MCAP
tags:
- robotics
- tactile
- humanoid
- manipulation
- ros2
- mcap
Lidar-Tactile — Raw ROS 2 MCAP recordings
Raw rosbag2 MCAP recordings from a Unitree G1 humanoid performing bimanual tactile
loco-manipulation with ViTac/sensx force sensors. These are the unprocessed bags;
the segmented training zarrs derived from them live in
GeorgiaTech/ik-modulation-zarr.
Layout
mcap/<task>/<stream>/<run>/<traj>/*.mcap (+ metadata.yaml per bag)
mcap/bucket/ — bucket manipulation task (2026-05)
| Stream | Path | Count |
|---|---|---|
| Teleop | mcap/bucket/teleop/2026-05-09-17-08-12/{001..056}/ |
56 trajectories |
| Human | mcap/bucket/human/{003..020}/ |
15 trajectories |
- Teleop: operator drives the robot; bags contain the commanded IK target
(
/g1_upper_pink_controller/{left,right}_hand), the measured EE (/tf_server_{left,right}/pose_measured), IK metrics, and tactile force/raw. - Human: human demonstration; measured hand poses + tactile force/raw.
Key topics
| Quantity | Topic |
|---|---|
| Measured EE (teleop) | /tf_server_{left,right}/pose_measured |
| IK target (teleop) | /g1_upper_pink_controller/{left,right}_hand |
| Tactile force | /sensor1/sensor1/force (left), /sensor2/sensor2/force (right), /chest_sensor/... |
| Tactile raw | /sensor{1,2}/.../raw, /chest_sensor/.../raw |
Sensor mapping: sensor1 = left hand, sensor2 = right hand.
Reading a bag
from mcap_ros2.decoder import DecoderFactory
from mcap.reader import make_reader
with open("mcap/bucket/teleop/2026-05-09-17-08-12/001/001_0.mcap", "rb") as f:
reader = make_reader(f, decoder_factories=[DecoderFactory()])
for schema, channel, message, ros_msg in reader.iter_decoded_messages():
... # channel.topic, ros_msg fields
See the conversion pipeline writeup (ik_modulation/configs/bucket_pipeline/BUCKET_PIPELINE.md
in the diffusion_humanoid_train repo) for how these bags become segmented training zarrs.