docker-ohm

CUDA + ROS 2 Humble container for CSIRO OHM — GPU-accelerated probabilistic occupancy / NDT-OM voxel mapping.

Subscribes to a sensor_msgs/PointCloud2 (from docker-depth-anything) plus TF (from docker-mocap4ros2-optitrack + docker-robot-setup-optitrack) and builds a live 3-D occupancy map.

Topics / services

Name	Type	Description
/depth_anything_v3/points	sensor_msgs/PointCloud2	Input cloud (subscribed)
/ohm/occupied_cloud	sensor_msgs/PointCloud2	Occupied voxel centres, map frame
/ohm/save	std_srvs/Trigger	Flush GPU→CPU and write .ohm + .ply

TF chain (dynamic map)

map
 └─ optitrack              ← docker-mocap4ros2-optitrack
     └─ cubr_rg_01
         └─ base_link
             └─ …
                 └─ irayple_rgb_optical_frame   ← point cloud frame

The map grows dynamically as the robot moves — no static TF needed.

Build & run

cd docker-ohm
docker compose build        # first time only (~20 min, downloads CUDA + OHM)
docker compose up -d

# inside the container — first time only:
source /opt/ros/humble/setup.bash
cd /root/user_ws
colcon build --symlink-install
source install/setup.bash

Subsequent shells auto-source the workspace via .bashrc.

Launch

# inside the container:
ros2 launch ohm_ros2 ohm_live.launch.py

# or directly from host:
docker exec -it ohm-ros2 bash -lc "ros2 launch ohm_ros2 ohm_live.launch.py"

Key launch arguments

ros2 launch ohm_ros2 ohm_live.launch.py \
    resolution:=0.1 \
    use_ndt:=true \
    range_min:=0.3 \
    range_max:=8.0

Param	Default	Notes
resolution	0.1	Voxel size in metres
range_min / range_max	0.3 / 8.0	Clip depth cloud (monocular depth is unreliable at extremes)
hit_probability / miss_probability	0.7 / 0.45	Log-odds update weights
use_ndt	true	NDT-OM mode; false = plain occupancy
ndt_sensor_noise_m	0.15	NDT sensor noise (monocular depth is noisier than stereo)
publish_rate_hz	2.0	Occupied-cloud republish rate
cloud_topic	/depth_anything_v3/points	Input point cloud topic

Save the live map

One service call writes both formats:

ros2 service call /ohm/save std_srvs/srv/Trigger {}

• /data/ohm_live.ohm — native OHM map
• /data/ohm_live.ply — voxel PLY for CloudCompare / MeshLab

The /data/ directory is bind-mounted from ../../bags/ on the host.

View in RViz2

Fixed Frame : map
Add → PointCloud2 → /ohm/occupied_cloud
Add → TF

Tuning for monocular depth

Monocular depth (Depth Anything V3) is noisier than stereo. Recommended starting values:

range_min: 0.3
range_max: 8.0
miss_probability: 0.45
use_ndt: true
ndt_sensor_noise_m: 0.15

Use the FLS mask (fls_mask:=true in docker-depth-anything) to restrict the input cloud to the sonar field of view — this removes out-of-region clutter before it reaches OHM.

Layout

docker-ohm/
├── Dockerfile               CUDA 12.x + Ubuntu 22.04 + OHM + ROS 2 Humble
├── docker-compose.yml
├── run.sh                   convenience wrapper (GPU + host net)
├── dds_profiles/            CycloneDDS / FastDDS / Zenoh profiles
└── user_ws/src/ohm_ros2/
    ├── config/ohm_live.yaml
    ├── launch/ohm_live.launch.py
    └── src/ohm_live_node.cpp

The user_ws/ tree is bind-mounted — edit source on the host and colcon build inside the container; no image rebuild needed.

DDS

Profile files are in dds_profiles/ (CycloneDDS loopback + large receive buffer).
Set RMW_IMPLEMENTATION via the environment or in run.sh. All other containers in this workspace use rmw_cyclonedds_cpp — this container must match.