RL Deploy - Go2 Parkour Deployment for ROS Noetic Gazebo (Testing version, bad perfomance)

Deploy RL parkour policies for Go2 quadruped robot in Gazebo simulation based on Extreme-Parkour-Onboard code (https://github.com/change-every/Extreme-Parkour-Onboard)

Overview

This package provides a complete deployment system for running trained RL policies on the Go2 quadruped robot in Gazebo. It includes:

• FSM Controller: Finite State Machine for managing robot states (Idle → Sit → Stand → RL Mode)
• Observation Manager: Aggregates sensor data and constructs observation vectors
• RL Policy Node: Executes trained policy inference with vision and proprioception
• Joystick Interface: Control robot states and velocity commands

Dependencies

# ROS Noetic
sudo apt install ros-noetic-desktop-full
sudo apt install ros-noetic-gazebo-ros-pkgs
sudo apt install ros-noetic-controller-manager
sudo apt install ros-noetic-joint-state-controller
sudo apt install ros-noetic-position-controllers
sudo apt install ros-noetic-joy

# Python packages
pip3 install torch torchvision
pip3 install opencv-python
pip3 install numpy

Trained Models

Place your trained models in ~/go2_parkour/traced/:

• base_jit.pt - Base policy (actor, estimator, history encoder)
• vision_weight.pt - Vision encoder weights

Usage

1. Launch Complete System

roslaunch rl_deploy complete_system.launch world:=parkour

This launches:

• Gazebo with Go2 robot
• All deployment nodes
• Joystick interface
• RViz (optional)

2. FSM Control Flow

State Machine

    IDLE (Robot spawned)
      │
      │ Press A
      ▼
    SIT (Sitting position)
      │
      │ Press B
      ▼
    STAND (Standing position)
      │
      │ Press Y
      ▼
    RL_MODE (Policy control active)

Configuration

Edit config/go2_parkour_config.yaml to customize:

Joint Positions

• joint_positions/sit: Sitting pose
• joint_positions/stand: Standing pose
• joint_positions/default: RL mode neutral pose

Control Parameters

control:
  static_mode:
    stiffness: 40.0  # PD gains for sit/stand
    damping: 2.0
  rl_mode:
    stiffness: 40.0  # PD gains for RL control
    damping: 1.0
  action_scale: 0.25
  control_dt: 0.02  # 50Hz

Observation Configuration

observation:
  n_proprio: 53
  n_depth_latent: 32
  history_len: 10
  scales:
    lin_vel: 2.0
    ang_vel: 0.25
    dof_pos: 1.0
    dof_vel: 0.05

Depth Camera

depth_camera:
  crop_top: 60
  crop_bottom: 100
  crop_left: 80
  crop_right: 36
  resize_width: 87
  resize_height: 58
  visual_update_interval: 5  # 10Hz

Observation Space

The system constructs a 123-dimensional observation vector matching the trained policy:

Component	Dimensions	Description
Angular velocity	3	Body frame gyroscope
IMU orientation	2	Roll, pitch
Yaw info	3	Delta yaw (from vision)
Commands	3	vx, vy, vyaw from joystick
Terrain flags	2	Parkour/walk mode
Joint positions	12	Relative to default
Joint velocities	12	Scaled
Last actions	12	Previous motor commands
Contact states	4	Foot contact flags
Depth latent	32	Vision-encoded features
Estimated lin vel	9	From estimator network
History latent	29	From history encoder
Total	123

Note: No height scanner is used (matching real hardware deployment).

Topics

Published

• /go2_joint_controller/command - Joint position commands
• /rl/commands - Velocity commands (vx, vy, vyaw)
• /rl/proprioception - Current proprioception vector
• /rl/last_action - Last action for observation
• /fsm/state - Current FSM state

Subscribed

• /joy - Joystick input
• /imu/data - IMU data
• /joint_states - Joint positions/velocities
• /camera/depth/image_raw - Depth camera
• /FR_foot_contact, /FL_foot_contact, etc. - Contact sensors

Based on the Isaac Lab extreme parkour training framework and real hardware deployment code.