RNL: Robot Navigation Learning Framework

End-to-end Deep Reinforcement Learning for Real-World Robotics Navigation in PyTorch

Overview

RNL (Robot Navigation Learning) is a comprehensive framework for training autonomous robots to navigate in unknown environments using Deep Reinforcement Learning (DRL). The framework provides a complete pipeline from environment simulation to model deployment, with support for both training and inference phases.

Demo

Key Features

• 3D environment
• Wandb integration
• LIDAR simulation
• Robot parameters
• Differential drive
• Physics simulation
• Dynamic maps
• LLM integration

Installation

Prerequisites

• Python 3.10+
• PyTorch 2.5.1+
• CUDA (optional, for GPU acceleration)

Install from PyPI

pip install rnl

Quick Start

1. Basic Training Example

import numpy as np
import rnl as vault

# Configure robot parameters
robot_config = vault.robot(
    base_radius=0.105,        # Robot radius in meters
    max_vel_linear=0.22,      # Maximum linear velocity
    max_vel_angular=2.84,     # Maximum angular velocity
    wheel_distance=0.16,      # Distance between wheels
    weight=1.0,               # Robot weight in kg
    threshold=1.0,            # Obstacle detection threshold
    collision=0.5,            # Collision radius
    path_model="None"         # Path to pretrained model
)

# Configure LIDAR sensor
sensor_config = vault.sensor(
    fov=2 * np.pi,           # Field of view (360 degrees)
    num_rays=20,             # Number of LIDAR rays
    min_range=0.0,           # Minimum detection range
    max_range=6.0            # Maximum detection range
)

# Configure environment
env_config = vault.make(
    scale=100,               # Environment scale
    folder_map="None",       # Custom map folder
    name_map="None",         # Custom map name
    max_timestep=10000       # Maximum episode length
)

# Configure rendering
render_config = vault.render(
    controller=False,       # Disable manual control (set True to control robot with arrow keys)
    debug=True,             # Enable debug visualization
    plot=False              # Disable plotting
)

# Initialize trainer
trainer = vault.Trainer(robot_config, sensor_config, env_config, render_config)

# Start training
trainer.learn(
    max_timestep_global=3000000,  # Total training steps
    seed=1,                       # Random seed
    batch_size=1024,             # Training batch size
    hidden_size=128,             # Neural network hidden size
    num_envs=4,                  # Parallel environments
    device="cuda",               # Training device
    checkpoint=10000,            # Checkpoint frequency
    use_wandb=True,             # Enable Wandb logging
    lr=0.0003,                  # Learning rate
    learn_step=512,             # Learning step size
    gae_lambda=0.95,            # GAE lambda parameter
    ent_coef=0.0,               # Entropy coefficient
    vf_coef=0.5,                # Value function coefficient
    max_grad_norm=0.5,          # Gradient clipping
    update_epochs=10,           # Update epochs
    name="navigation_model"      # Model name wandb
)

2. Inference Example

import rnl as vault

# Use same configuration as training
# ... (robot_config, sensor_config, env_config, render_config)

# Initialize simulation
simulation = vault.Simulation(robot_config, sensor_config, env_config, render_config)

# Run inference
simulation.run()

3. Demo Mode

python main.py -m sim

Advanced Features

LLM Integration

The framework supports Large Language Model integration for automated reward engineering:

trainer.learn(
    use_llm=True,
    llm_api_key="your_api_key",
    population=2,
    loop_feedback=10,
    description_task="reach the goal while avoiding obstacles efficiently"
)

Parallel Training

The framework supports multi-environment parallel training for faster convergence:

trainer.learn(
    num_envs=8,              # Number of parallel environments
    device="cuda",           # Use GPU acceleration
    batch_size=2048          # Larger batch size for parallel training
)

Contributing

We welcome contributions! Please see our Contributing Guide for details.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Support

• Issues: GitHub Issues
• Email: grottimeireles@gmail.com