Quadruped robot dog simulation, walking control, and reinforcement learning policy training workspace.
Supports: Unitree Go1/Go2, Boston Dynamics Spot, MIT Mini Cheetah, ANYmal B/C, Mini Pupper.
quadruped-dog-rl/
├── urdf/ # Robot URDF and mesh files
│ ├── go1_config/ # Unitree Go1
│ ├── go2_unitree/ # Unitree Go2 (with DAE meshes)
│ ├── spot_config/ # Boston Dynamics Spot
│ ├── mini_cheetah_config/ # MIT Mini Cheetah
│ ├── mini_pupper_config/ # Mini Pupper
│ ├── anymal_b_config/ # ANYmal B (ETH Zurich)
│ └── anymal_c_config/ # ANYmal C (ETH Zurich)
├── ros2/ # ROS2 packages (CHAMP framework, ros2 branch)
│ ├── champ/ # Core locomotion controller
│ ├── champ_base/ # Hardware abstraction layer
│ ├── champ_bringup/ # Launch files
│ ├── champ_config/ # Robot-specific configs
│ ├── champ_description/ # URDF loading
│ ├── champ_gazebo/ # Gazebo simulation
│ ├── champ_navigation/ # Navigation stack
│ ├── champ_teleop/ # Keyboard/joystick teleoperation
│ └── robots/ # Pre-configured robot packages
├── launch/ # Top-level launch files
│ ├── view_go2.launch.py # View Go2 URDF in RViz2
│ ├── gazebo_go2.launch.py # Spawn Go2 in Gazebo Garden
│ ├── gazebo_sim.launch.py # Generic Gazebo sim launcher
│ ├── rviz_view.launch.py # Generic RViz2 viewer
│ └── policy_deploy.launch.py # Deploy trained RL policy
├── scripts/ # Shell scripts for common tasks
│ ├── train_policy.sh # Train walking policy
│ ├── play_policy.sh # Visualize trained policy
│ └── launch_sim.sh # Launch Gazebo sim
├── training/ # RL policy training (Unitree RL Gym)
│ ├── legged_gym/ # PPO training scripts and environments
│ ├── deploy/ # Policy deployment to real robot
│ └── setup.py
├── description/ # Robot description docs and joint conventions
└── interfaces/ # Custom ROS2 msgs, srvs, actions (placeholder)
- Ubuntu 22.04
- ROS2 Humble
- Gazebo Garden (gz-sim7) — already works with
ros_gz_sim - Python 3.8+
- NVIDIA GPU with 10GB+ VRAM for RL training
cd ros2
source /opt/ros/humble/setup.bash
colcon build --symlink-install --cmake-args -DBUILD_TESTING=OFF
source install/setup.bashsource /opt/ros/humble/setup.bash
ros2 launch launch/view_go2.launch.pyOpens RViz2 with the full Go2 mesh and a joint slider GUI to pose the legs.
source /opt/ros/humble/setup.bash
ros2 launch launch/gazebo_go2.launch.pyStarts Gazebo Garden, spawns the Go2, bridges topics to ROS2, and opens RViz2 alongside it.
Send velocity commands:
ros2 topic pub /cmd_vel geometry_msgs/msg/Twist "{linear: {x: 0.5}}" --onceRequires Isaac Gym — download from https://developer.nvidia.com/isaac-gym
cd training
pip install -e .
# Train Go2 walking policy
python legged_gym/scripts/train.py --task=go2 --headless
# Visualize trained policy
python legged_gym/scripts/play.py --task=go2
# Deploy to real robot
cd deploy
python deploy.py --task=go2 --ckpt=<path_to_checkpoint>Or use the helper scripts:
./scripts/train_policy.sh go2
./scripts/play_policy.sh go2| Robot | URDF Path | RL Task |
|---|---|---|
| Unitree Go1 | urdf/go1_config/ |
go1 |
| Unitree Go2 | urdf/go2_unitree/urdf/go2.urdf |
go2 |
| Boston Dynamics Spot | urdf/spot_config/ |
— |
| MIT Mini Cheetah | urdf/mini_cheetah_config/ |
— |
| ANYmal B | urdf/anymal_b_config/ |
— |
| ANYmal C | urdf/anymal_c_config/ |
— |
| Mini Pupper | urdf/mini_pupper_config/ |
— |
Higher-level autonomy stack built on top of the base simulation and RL policy.
intelligence/
├── gait/
│ └── gait_scheduler.py # Auto-select gait (walk/trot/canter/bound) by speed
├── perception/
│ └── terrain_estimator.py # Classify terrain (flat/slope/stairs/rough) from IMU + foot forces
├── navigation/
│ └── waypoint_navigator.py # Autonomous waypoint following via pure pursuit
├── terrain/
│ └── adaptive_controller.py # Fuse terrain + gait into safe velocity commands
└── llm_commander/
└── llm_commander.py # Natural language -> robot commands via Claude API
Auto-selects the right gait based on commanded speed:
| Speed (m/s) | Gait | Foot pattern |
|---|---|---|
| 0 – 0.05 | Stand | All feet down |
| 0.05 – 0.4 | Walk | One foot at a time |
| 0.4 – 1.5 | Trot | Diagonal pairs (FL+RR, FR+RL) |
| 1.5 – 2.5 | Canter | Three-beat |
| 2.5 – 4.0 | Bound | Front pair then rear pair |
| 4.0+ | Pronk | All four feet airborne |
Classifies terrain from IMU and foot contact forces, outputs recommended speed limit and foot clearance:
from intelligence.perception.terrain_estimator import TerrainEstimator
estimator = TerrainEstimator()
result = estimator.estimate(imu_roll=0.1, imu_pitch=0.05, contacts=[120, 115, 118, 122])
# TerrainEstimate(terrain_type=flat, slope_deg=6.6, recommended_speed_limit=3.0)Autonomous point-to-point navigation using pure pursuit:
ros2 run quadruped_dog_rl waypoint_navigator --ros-args \
-p waypoints:="[[2.0,0.0],[2.0,2.0],[0.0,2.0],[0.0,0.0]]" \
-p linear_speed:=0.5Control the robot with plain English using Claude API:
export ANTHROPIC_API_KEY=your_key
python3 intelligence/llm_commander/llm_commander.pyThen publish commands:
ros2 topic pub /natural_language_cmd std_msgs/msg/String "data: 'trot forward at medium speed'"
ros2 topic pub /natural_language_cmd std_msgs/msg/String "data: 'turn left slowly'"
ros2 topic pub /natural_language_cmd std_msgs/msg/String "data: 'stop'"Combines terrain estimation + gait scheduling into a single safe command output:
from intelligence.terrain.adaptive_controller import AdaptiveController
ctrl = AdaptiveController()
cmd = ctrl.adapt(desired_speed=1.2, imu_pitch=0.12, contacts=[110,115,108,120])
# AdaptedCommand(linear_x=1.0, gait='trot', terrain='slope', foot_clearance=0.08)- CHAMP Framework — ROS2 locomotion controller
- Unitree RL Gym — PPO policy training
- legged_gym (ETH Zurich) — original RL gym
- Isaac Lab — modern GPU training framework