feat: Dockerized dev environment, Gazebo-MuJoCo co-simulation fixes, Nav2 stack repair, and rl_sar_zoo as submodule

[hugodidi]

Summary

This PR adds a complete containerized development setup, fixes simulation stability and visual fidelity,
and restructures rl_sar_zoo as a proper git submodule. All changes target the Unitree G1 29-DoF humanoid
navigating via Nav2 in a Gazebo + MuJoCo co-simulation.

Key changes

Docker environment (docker/)

• Add Dockerfile + docker-compose.yml for ROS 2 Humble with MuJoCo, Gazebo Classic, and Nav2
• Install Eclipse CycloneDDS for reliable inter-process communication
• Container layout matches README instructions for reproducibility
• Add Docker section to README with build/enter instructions

Gazebo ↔ MuJoCo segment-to-segment synchronization (gazebo_mujoco_pose_sync)

• Refactor plugin to use gazebo_ros::Node (was raw Gazebo plugin), fixing multithreading segfaults that crashed gzserver
• Implement segment-to-segment (link-by-link) pose synchronization of the G1 from MuJoCo to Gazebo
• Gazebo no longer participates in physics simulation — the robot is kinematically driven by MuJoCo poses, so it passes through obstacles in Gazebo. This eliminates collision solver errors that caused gzserver crashes. A future system is planned to relay obstacle information from Gazebo/sensors back to MuJoCo for physics-aware collision handling
• Fix collision configuration during pose transfer

Texture and visual fixes

• Fix texture upload paths in turtlebot3_house model (model.sdf)
• Inline sun and ground_plane models into turtlebot3_house.world (required because GAZEBO_MODEL_DATABASE_URI="" in spawn_g1.launch.py blocks online model downloads)

rl_sar modifications (src/rl_sar/)

• Fix joint state publishing: publish /mujoco/joint_states as sensor_msgs/JointState
• Fix joint mapping in fsm_g1.hpp: use hardware mapping (not policy mapping) during locomotion
• Add missing #include in 8 unitree_sdk2/log/*.hpp headers
• Scope .gitignore rules (was ignoring files outside rl_sar)
• Fix build.sh script path references

G1 description & Gazebo launch (g1_description, g1_gazebo)

• Set g1_29dof as default simulation model (remove g1_23dof.urdf)
• Adjust Gazebo plugin z_offset to -0.82 for correct ground alignment
• Restore LiDAR to original config: 1440 samples, 20 Hz, horizontal orientation

Nav2 stack (g1_nav2, humanoid_nav_bridge)

• Repair nav_amcl.launch.py configuration for AMCL + costmaps
• Add pre-built map files (g1_map.pgm, g1_map.yaml)
• Add mujoco_to_gazebo_node.py bridge node

rl_sar_zoo as git submodule

• Register src/rl_sar/src/rl_sar_zoo as submodule → fan-ziqi/rl_sar_zoo (was downloaded at build time by build.sh)
• Remove rl_sar_zoo from src/rl_sar/.gitignore (now tracked as submodule)
• Update README: add COLCON_IGNORE touch and package.xml symlink steps after clone

How to test

# Clone
git clone --recurse-submodules https://github.com/uleroboticsgroup/rl_hnav.git
cd rl_hnav
touch src/rl_sar/src/rl_sar_zoo/COLCON_IGNORE
for d in src/rl_sar/src/robot_msgs src/rl_sar/src/robot_joint_controller src/rl_sar/src/rl_sar; do
  ln -sf package.ros2.xml "$d/package.xml"
done

# Build (inside container)
xhost +local:docker
docker compose -f docker/docker-compose.yml up -d --build
docker exec -it rl_hnav_container bash
source /opt/ros/humble/setup.bash
cd ~/rl_hnav && colcon build --cmake-args -DUSE_MUJOCO=ON -DENABLE_REAL_ROBOT=ON
source install/setup.bash

# Launch Nav2
ros2 launch g1_nav2 nav_amcl.launch.py
# In another terminal: launch MuJoCo locomotion
ros2 run rl_sar rl_mujoco g1 scene_29dof --ros-args -p navigation_mode:=true

Architecture note:

Gazebo acts as a sensor server (LiDAR, camera) and visualizer only. MuJoCo owns all physics.
Obstacle information currently flows one-way (Gazebo sensors → Nav2 → MuJoCo via /cmd_vel).
Direct obstacle-to-MuJoCo relay for contact-aware locomotion is planned as future work.

2026-03-25.23-59-50.mp4