IsaacLegs

A reinforcement-learning and deployment pipeline for legged robots, built on NVIDIA Isaac Lab, Isaac Sim, and ROS 2.

IsaacLegs facilitates the deployment of policies trained in Isaac Lab onto robots inside the Isaac Sim digital twin using ROS. It currently supports the Unitree Go2 quadruped and the Unitree G1 humanoid, including a 29-DOF configuration, with additional robots coming soon.

The framework provides ready-to-use digital twins that integrate robot state, sensor streams, and control interfaces within Isaac Sim, with support for creating custom digital twins for additional robot platforms. Robots can be deployed into a variety of NVIDIA simulation environments, referenced from Isaac Asset or Custom Environments.

The controller is designed to work with different robot platforms through ROS 2 interfaces, making it easy to support additional robots, sensors, and policies.

---

Guides

• How to train a policy — train in Isaac Lab and export the deployable artifacts.
• How to create a digital twin — import the robot, set drive gains, wire the ROS 2 action graphs.
• How to use the policy controller for a custom robot — use general controller to deploy your policy.

Roadmap

• Locomotion policy training in Isaac Lab (Go2, G1, G1 29-DOF) ✅
• Ready-to-use ROS 2 digital twins in Isaac Sim ✅
• Robot-agnostic policy controller ✅
• Onboard sensor integration (camera, 3D LiDAR) ✅
• Deployment into stock NVIDIA simulation environments ✅
• RViz configuration for all robots
• SLAM and navigation
• Imitation learning projects

---

Gallery

Policies trained in Isaac Lab, and deployed in Isaac Sim digital twins across different environments.

Go2 — Warehouse	G1 — Flat ground
G1 (29-DOF) — Simple room	Go2 — Camera + 3D LiDAR in RViz

---

Requirements

Component	Version	Notes
Ubuntu	22.04
Isaac Sim	≥ 5.0.0	the simulator and digital-twin runtime
Isaac Lab	≥ 0.46.2	needed only to train policies — its pip install also pulls in Isaac Sim
ROS 2	Humble
Python	3.10

A CUDA-capable NVIDIA GPU is required for Isaac Sim.

---

Getting started

A complete run consists of the digital twin in Isaac Sim and the policy controller in ROS 2. The steps below take you from a fresh clone to a walking robot.

1. Clone and build the workspace

git clone https://github.com/gholibqasobov/IsaacLegs.git ~/IsaacLegs
cd ~/IsaacLegs

# Install ROS 2 package dependencies 
rosdep update
rosdep install --from-paths src --ignore-src -r -y

# The policy controller runs a TorchScript policy
pip install torch

colcon build --symlink-install
source install/setup.bash          # re-run in every new shell

2. Launch a digital twin

launch_scene.py opens the Isaac Sim GUI with a ready-to-run robot and environment. Run the script from the Isaac Sim Python environment using the option that matches your installation:

If you have a conda with IsaacSim/IaacLab installed:

conda activate env_isaaclab
cd ~/IsaacLegs
python isaacsim_envs/gui/launch_scene.py --go2 --warehouse --play

Or use Isaac Sim's bundled interpreter (python.sh):

cd ~/IsaacLegs
~/isaacsim/python.sh isaacsim_envs/gui/launch_scene.py --go2 --warehouse --play

Adjust ~/isaacsim to match your Isaac Sim installation path. The same approach applies to all other Isaac Sim scripts.

Options. Pick one robot, optionally one environment, plus any run flags — e.g. --go2 --warehouse --play.

• Robot: --go2 · --g1 · --g1_29dof
• Environment: --flat_grid · --flat_plane · --rough_plane · --simple_room · --warehouse · --full_warehouse
• --play — start simulating immediately; Otherwise press ▶ Play in the GUI.
• --headless — run without a GUI window.
• --list — print the available scenes and environments, then exit.

See Supported robots & environments for previews and details.

3. Run the policy controller

In a second terminal (with the workspace sourced):

ros2 launch fullbody_controller go2.launch.py     # also: g1.launch.py, g1_29dof.launch.py

Start order matters for humanoids. A G1 will fall in the moments before the policy engages. Launch the scene without --play, start the controller, then press ▶ Play — or start the controller first and launch with --play. Quadrupeds (Go2) are forgiving either way.

Parameters (-p name:=value on ros2 run, or name:=value on ros2 launch):

• policy_path — path to the TorchScript policy (.pt); defaults to the shipped checkpoint.
• io_descriptors_path — override the descriptor location (default: next to the policy).
• decimation (default 1) — run the policy every Nth tick (50 Hz sensors → 50 Hz control at 1).
• warmup_sec (default 0.0) — seconds spent easing into the policy's default pose before engaging it.
• warmup_interpolate (default True) — interpolate from the spawn pose to default (no violent snap).

4. Command and inspect the robot

Drive it by publishing a velocity, or interactively from the keyboard:

ros2 run teleop_twist_keyboard teleop_twist_keyboard
# or
ros2 topic pub /cmd_vel geometry_msgs/msg/Twist "{linear: {x: 0.3}, angular: {z: 0.3}}" -r 10

Inspect what's flowing over ROS 2:

ros2 topic list
ros2 topic echo /joint_states

---

Supported robots & environments

Currently Available Robots

Preview	Robot	Flag	Policy directory
	Unitree Go2 — quadruped	--go2	policy/go2_locomotion
	Unitree G1 — humanoid	--g1	policy/g1_locomotion
	Unitree G1 29-DOF — humanoid	--g1_29dof	policy/g1_29dof_locomotion

Policy directories are under src/fullbody_controller/policy/

Sample Stock Environments

Preview	Environment	Flag
	Flat grid	--flat_grid
	Flat plane	--flat_plane
	Simple room	--simple_room
	Warehouse	--warehouse
	Full warehouse	--full_warehouse

Adding a stock world is a one-line edit to the ENVIRONMENTS dict in launch_scene.py:

ENVIRONMENTS = {
    # ...
    "my_world": "/Isaac/Environments/.../my_world.usd",
}

ROS 2 topics.

Published by the simulation (action graphs):

Topic	Type
joint_states	sensor_msgs/JointState
joint_command	sensor_msgs/JointState
imu	sensor_msgs/Imu
odom	nav_msgs/Odometry
clock	rosgraph_msgs/Clock
/rgb, /depth	sensor_msgs/Image
/scan, /point_cloud	sensor_msgs/{LaserScan,PointCloud2}

---

Contributing & feedback

If IsaacLegs is useful to you, please consider giving it a ⭐ — it helps others find the project and motivates further development. Found a bug, have a feature in mind, or want a new robot or environment supported? Open an issue or start a discussion — feedback and contributions are very welcome.

---

License

Apache-2.0. See individual package headers.

Acknowledgements

NVIDIA Isaac Lab and Isaac Sim, Unitree Robotics, and the RSL-RL project.