Isaac Sim Quadruped (Trakr)

This repository contains the necessary files and scripts to run a trakr quadruped robot simulation in NVIDIA Isaac Sim. The project is configured for reinforcement learning-based control, featuring environment setup, sensor simulation, and a client-server architecture for data streaming and remote control.

Table of Contents

• Isaac Sim Quadruped (Trakr)
  • Table of Contents
  • 1. Getting Started
    • Installation
  • 2. Project Structure
    • Breakdown
  • 3. Configuration Deep Dive
    • env.yaml
    • run_params.yaml
    • tcp_config.yaml
  • 4. Running the Simulation
  • 5. Robot and Policy
  • 6. Troubleshooting

---

1. Getting Started

Installation

1. Download and install Isaac Sim Standalone:

wget https://download.isaacsim.omniverse.nvidia.com/isaac-sim-standalone%404.5.0-rc.36%2Brelease.19112.f59b3005.gl.linux-x86_64.release.zip
mkdir ~/isaacsim
unzip "isaac-sim-standalone@4.5.0-rc.36+release.19112.f59b3005.gl.linux-x86_64.release.zip" -d ~/isaacsim

2. Clone this repository inside the Isaac Sim installation path:

cd ~/isaacsim/isaac-sim-standalone@4.5.0-rc.36+release.19112.f59b3005.gl.linux-x86_64.release/
git clone https://github.com/HumanoidAddverb/trakr_slam_sdk.git

2. Project Structure

The project is organized into four main directories:

.
├── config/             # YAML files for simulation, environment, and network settings.
├── models/             # Contains the pre-trained neural network policy.
├── resources/          # Robot 3D models (USD, URDF) and meshes (STL).
└── scripts/            # Python scripts to run the simulation and client/server logic.

Breakdown

• config/:
  • env.yaml: Defines the RL environment, including physics, observations, rewards, and terminations.
  • run_params.yaml: High-level simulation parameters, asset paths, and sensor configurations.
  • tcp_config.yaml: Network settings for the client-server communication.
• models/:
  • policy.pt: The trained PyTorch model that controls the quadruped's locomotion.
• resources/:
  • meshes/: Contains all the individual .STL files for the robot's links.
  • robot_description/: Contains the robot's full description files, including trakr.urdf and the necessary USD files (trakr_imu.usd, full_warehouse.usd) for Isaac Sim.
• scripts/:
  • trakr_standalone.py: The main entry point. It launches the Isaac Sim application, sets up the environment, and runs the simulation. It also acts as the TCP server.
  • trakr.py: Defines the TrakrFlatTerrainPolicy class, which loads the RL policy and computes joint torques based on observations.
  • server_data_transfer.py: A simple TCP server class for sending data.
  • client_data_transfer.py: A TCP client that receives and visualizes sensor data, and sends keyboard commands back to the simulation.

3. Configuration Deep Dive

env.yaml

This is the configuration file, defining the reinforcement learning environment for training the policy.

• sim: Contains core physics settings, including the physics dt (0.002s), gravity, and detailed PhysX engine parameters.
• scene: Defines all assets in the simulation.
  • robot: Specifies the robot's USD path, initial state (position, rotation, joint angles), actuator properties (PD controller gains), and collision settings.
  • terrain: Configures the ground plane. It can generate various curricula of terrains, from flat to random rough patches.
  • contact_forces: Sets up a sensor to detect contacts on the robot's body, used for termination conditions.
• observations: Defines the observation vector fed to the policy network. It includes:
  • Base linear and angular velocity.
  • Projected gravity vector.
  • Velocity commands.
  • Joint positions and velocities.
  • The previous action taken.
• rewards: A collection of weighted reward functions that guided the policy's training. These include rewarding desired linear/angular velocity, proper gait, and foot clearance, while penalizing high torques, base motion, and unwanted contact.
• terminations: Conditions that end an episode, such as the robot falling over (body_contact), going out of bounds, or timing out.
• commands: Defines how target velocity commands are generated for the robot to follow.

run_params.yaml

This file provides high-level parameters and asset paths used by the simulation scripts.

• Control Parameters: action_scale, p_gain, d_gain, and default_joint_pos for the policy and joint controller.
• Asset Paths: Specifies the exact paths to the environment_usd_path and trakr_usd_path. You may need to edit these paths if your project is not in the default location.
• Spawning Info: Defines the robot's initial trakr_spawn_position and trakr_spawn_orientation.
• Sensor Intrinsics: Contains camera parameters (K, D), resolution (width, height), and sensor-to-base_link transforms.

tcp_config.yaml

This file configures the network settings for the client-server communication.

• data_server / data_client: Sets the IP and port for the primary data stream (sensor data from sim to client, commands from client to sim).
• teleop_server / teleop_client: Configures a secondary connection, likely intended for a separate teleoperation interface.

4. Running the Simulation

The entire project can be run using the trakr_standalone.py script, which launches the Isaac Sim environment and the data server. For visualization and control, you will then run the client script.

1. Launch the Simulation

Navigate to the scripts directory in your terminal and execute:

cd ~/isaacsim/isaac-sim-standalone@4.5.0-rc.36+release.19112.f59b3005.gl.linux-x86_64.release/trakr_simulator/scripts/

2. Update the Paths
Before running the simulation, make sure to edit the following paths in your setup:

• In config/run_params.yaml, update these entries to match your local installation path:

  trakr_usd_path: "/home/<username>/isaacsim/.../trakr_simulator/resources/robot_description/trakr_imu.usd"
  policy_path:   "/home/<username>/isaacsim/.../trakr_simulator/models/policy.pt"
  env_config_path: "/home/<username>/isaacsim/.../trakr_simulator/config/env.yaml"

• In scripts/trakr_standalone.py, update the dataset output path:

  base_dir = "/home/<username>/your_workspace/dataset"

3. Launch the Simulation

Open a second terminal, navigate to the scripts directory, and run:

../../python.sh trakr_standalone.py

4. Controlling the Robot
Once the simulation is running, open a new terminal and start the TCP client to connect and send control commands to the robot. bash

cd ~/isaacsim/isaac-sim-standalone@4.5.0-rc.36+release.19112.f59b3005.gl.linux-x86_64.release/trakr_simulator/scripts/

../../python.sh client_data_transfer.py

This will:
Receive RGB, Depth, LiDAR, and IMU data streams from the server.
Display a 2×2 visualization panel (left camera, right camera, depth map, LiDAR scatter).
Save IMU data to imu_data.csv

5. Robot and Policy

• Robot Model: The trakr robot is defined by the trakr.urdf file for its kinematics and dynamics, and the trakr_imu.usd file is used to represent it visually and physically within Isaac Sim.
• Control Policy: The robot's locomotion is not manually coded. Instead, it is driven by the neural network in models/policy.pt. This network was trained using reinforcement learning to map sensor observations to low-level joint actions to achieve stable walking and follow velocity commands.

6. Troubleshooting

• Path Errors: If the script fails to find a USD or model file, double-check the paths in run_params.yaml. They may need to be absolute paths depending on your setup.
• TCP Connection Fails:
  • Ensure the server (trakr_standalone.py) is running before you start the client.
  • Verify that the IP addresses in tcp_config.yaml are correct. Use 127.0.0.1 if running on the same machine. If on different machines, use the server's network IP and ensure no firewalls are blocking the port.
• Client Visuals Don't Appear: Make sure you have a desktop environment and that opencv-python and matplotlib are installed correctly.
• Slow Performance: This simulation is computationally intensive. If performance is poor, you can try:
  • Running Isaac Sim in headless mode.
  • Reducing the rendering resolution in run_params.yaml.
  • Disabling advanced rendering features in env.yaml.
• Robot is Unstable: The policy's performance is tied to the physics parameters in env.yaml and run_params.yaml. Changes to p_gain, d_gain, physics_dt, or mass properties can affect stability.