Generalized Open Layered Embodied Modules — the software stack for the Correll Lab's Unitree humanoids: the H1-2 and, beside it, the G1 (see The G1).
Two simulators (MuJoCo/RoboCasa and Isaac Lab) and a ROS 2 Humble workspace run
in separate Docker containers sharing one DDS domain. The workspace holds the
drivers, upper-body IK, lower-body RL locomotion, a safety layer, SLAM and
navigation, perception model servers, and an LLM-callable skills layer. Sim and
real robot speak the same Unitree rt/lowcmd / rt/lowstate wire format, so
code moves between them unchanged.
The stack runs on x86 Linux with an NVIDIA GPU. Full detail in
docs/SETUP.md and docs/RUNNING.md.
git lfs install && git submodule update --init --recursive
cp docker/.env.example docker/.env # set GEMINI_API_KEY and ROS_DOMAIN_ID (never 0)
docker/scripts/docker_build.sh robocasa rosThen, in two terminals:
docker/scripts/docker_run.sh robocasa # 1. simulator (start first)
docker/scripts/docker_run.sh ros # 2. workspace shell
ros2 launch golem_bringup h12_sim.launch.py # bringup is a manual stepThe sim runs the real robot's operator gates. The legs hold their pre-pose until you confirm them, and ENTER in terminal 1 (or in the sim's viewer) presses the estop; after an estop, restart the whole bringup. In a third terminal, confirm the legs and ask the robot to do something:
docker exec -it golem_ros bash
source /opt/ros/humble/setup.bash && source /home/code/core_ws/install/setup.bash
ros2 service call /lowerbody/confirm_engage std_srvs/srv/Trigger
ros2 action send_goal /skill/grasp custom_ros_messages/action/SkillGrasp \
"{target_object: 'vertical fridge handle', arm: 'right', timeout: {sec: 60, nanosec: 0}}" --feedbackThe workspace is core_ws/src. Both robots run the same golem_* packages,
each configured per robot.
| Package | What it does |
|---|---|
golem_bringup |
Launch files that start the whole robot, sim or real (CLAUDE.md §6) |
golem_safety |
The only writer of the motor command topic: relays, clips and checks every command, latches the estop |
golem_upper_ik |
Arm IK controller (/frame_task, /named_config) with planning and self-collision avoidance |
golem_lowerbody_rl |
Leg controller running the walking policies (ALMI, FAME and walk on the H1-2; HOMIE on the G1) |
golem_slam |
FAST-LIO odometry, slam_toolbox mapping and Nav2 navigation |
golem_skills |
The /skill/* actions an LLM can call (grasp, pick and place, frontier exploration) |
golem_ros2_models |
Both robots' URDFs, meshes and self-collision models for ROS |
model_server |
Gemini, SAM3, GraspGenX and YOLO behind ROS services |
lmg_deploy |
The loco-manipulation game's whole-body MARL policy |
dex_3 |
The G1's Dex3-1 hands |
Each package's README.md documents it; CLAUDE.md §4 lists the
vendored and submodule packages beside them. The RoboCasa sim for both robots is
golem_robocasa/, the Isaac Lab sim (H1-2) CL_isaaclab_sim/.
GOLEM also runs the Unitree G1 (29 DoF, two Dex3-1 hands): its own package
(dex_3, the hands), the packages it shares with the H1-2 above, its models in
CL_Assets/g1, and a RoboCasa sim in golem_robocasa/g1/. The H1-2 stays the
default; --g1 on a wrapper, or GOLEM_ROBOT=g1, selects the G1 (RoboCasa
only). The images are the same, and every ros container builds and runs both
robots' packages.
docker/scripts/docker_stack.sh up robocasa --g1 # the G1 sim and its bringup (the arms and HOMIE)
tests/g1/test_sim_bringup.sh # the G1 acceptance test (arms); see tests/g1/README.mdOn the real G1 the whole stack runs in Docker on a laptop wired to the robot,
with the robot in Unitree's debug mode. Nothing has run on it yet; the first
session follows docs/g1/first-session.md. Start
at docs/g1/index.md.
Licence note. HOMIE's walking policy in
core_ws/src/golem_lowerbody_rl/policies/homie/ (from
OpenHomie) is for
non-commercial use only. Its LICENSE is CC BY-NC 4.0, but OpenHomie's README
says CC BY-NC-SA 4.0, which also requires changed versions to be shared under
the same licence: treat it as CC BY-NC-SA 4.0. The packages around it
(golem_lowerbody_rl, which runs it, and dex_3) are MIT. See the policy's
README.md.
The closed-loop test runs a robot's whole stack in the RoboCasa sim: it stands
unsupported, walks to a Nav2 goal while moving its arms, keeps SLAM healthy, and
golem_safety stays the only /lowcmd writer with no estop.
tests/sim/test_sim_stack.sh --h12
tests/sim/test_sim_stack.sh --g1tests/README.md covers the rest: the sim unit tests, the
G1 acceptance tests and the Docker script tests. Each package's own tests run
with colcon test in the ros container.
The repo layout is in CLAUDE.md §2 (most of the tree is git
submodules, and large assets are Git-LFS). The user pages are in
docs/, and docker/BUILD.md covers how
the images are layered and what is pinned.