PPO LiDAR Navigation

Hybrid Deep RL + Nav2 indoor robot navigation — 82.7% success rate over 150 benchmark scenarios

---

🎥 Watch full demo recording (YouTube) 📁 For more detail analysis of output prefer this link 📊 Full results & failure analysis
🏗️ System architecture deep-dive

---

What This Is

A trained robot navigation system that combines Nav2 global path planning with a PPO (Proximal Policy Optimisation) deep RL agent for local navigation and reactive obstacle avoidance.

The robot navigates a 16 × 16 m simulated apartment with rooms, corridors, and furniture — using only a 360° LiDAR sensor and the Nav2 global path. No camera. No object detection. Pure sensor-to-action end-to-end control.

You can click anywhere in RViz and the robot will navigate there.

---

System Architecture

User clicks in RViz
        │
        ▼ /goal_pose
Nav2 Global Planner (SLAM map)
        │
        ▼ /plan  (global path)
┌───────────────────────────────────┐
│   PPO Actor  (50D → 2D)           │
│                                   │
│  36D  LiDAR sectors               │
│   2D  Goal waypoint               │
│   2D  Current velocity            │
│  10D  Nav2 path features  ──────► │──► /cmd_vel_nav ──► Robot
└───────────────────────────────────┘

The PPO agent sees a 50-dimensional observation and outputs smooth velocity commands at 10 Hz, handling tight turns, wall-blocked corridors, and complex obstacle clusters that trip up classical planners alone.

---

Results

150 scenarios across 8 categories — open spaces, wall-blocked corridors, complex obstacle clusters, and random start/goal pairs. Zero oscillations. Zero fallback triggers.

→ Full breakdown by category

---

Quick Demo (Mac)

Prerequisites

• macOS with Apple Silicon or Intel
• ~8 GB free disk space (ROS2 + dependencies)

1. Clone & Setup

git clone https://github.com/Sourav29-2/ppo-lidar-navigation-.git
cd ppo-lidar-navigation-
bash setup/setup_mac.sh

This installs everything: ROS2 Humble, Nav2, SLAM Toolbox, PyTorch, Gymnasium. Takes 5–15 minutes on first run.

2. Launch

./demo.sh

Gazebo, SLAM, Nav2, and RViz open automatically.

3. Navigate

1. In RViz, click the "Nav2 Goal" button in the toolbar (arrow icon)
2. Click anywhere on the apartment map
3. Watch the robot navigate there!
4. Click again to set a new destination anytime

---

Quick Demo (Windows)

Step 1 — Install WSL2 (PowerShell, as Administrator)

wsl --install -d Ubuntu-22.04

Restart when prompted.

Step 2 — Inside the Ubuntu WSL2 terminal

curl -fsSL https://raw.githubusercontent.com/Sourav29-2/ppo-lidar-navigation-/master/setup/setup_wsl_inner.sh | bash

Step 3 — Run

cd ~/ppo-lidar-navigation-
./demo.sh

→ Full Windows setup guide

---

Repository Structure

ppo-lidar-navigation-/
│
├── demo.py                  # Interactive RViz demo (inference only)
├── demo.sh                  # One-command launcher
│
├── ppo/                     # PPO algorithm implementation
│   ├── actor.py             # Policy network (Gaussian, 50D → 2D)
│   ├── critic.py            # Value network
│   ├── ppo_trainer.py       # PPO update step
│   └── rollout_buffer.py    # Experience collection
│
├── src/urdf_test/           # ROS2 package
│   ├── scripts/
│   │   └── gazebo_nav_env.py  # Gymnasium environment (GazeboMacNavEnv)
│   ├── launch/              # Gazebo + Nav2 + RViz launch files
│   ├── config/
│   │   ├── nav2_params.yaml
│   │   └── rviz_demo.rviz   # Pre-configured RViz layout
│   ├── maps/                # SLAM-generated apartment map
│   └── urdf/                # Robot URDF model
│
├── training/                # Training scripts
│   ├── train_phase5_safety.py
│   └── run_phase5_training.sh
│
├── evaluation/              # Evaluation scripts
│   ├── eval_final_phase5.py
│   └── run_final_eval.sh
│
├── checkpoints/
│   └── hybrid_phase5/
│       └── best_success.pt  # Final champion model
│
├── results/                 # Clean evaluation data
│   ├── final_evaluation.csv
│   ├── final_evaluation_report.txt
│   └── benchmark_150_scenarios.json
│
├── setup/                   # Platform setup scripts
│   ├── setup_mac.sh
│   ├── setup_windows_wsl.md
│   └── setup_wsl_inner.sh
│
├── ARCHITECTURE.md          # System design deep-dive
├── RESULTS.md               # Full benchmark results
├── ROADMAP.md               # Planned extensions
└── CONTRIBUTING.md          # How to extend the project

---

Training From Scratch

If you want to retrain the agent:

# Install dependencies
bash setup/setup_mac.sh

# Start simulation (in your ros2_study workspace)
# Then in a second terminal:
pixi run python training/train_phase5_safety.py

Training logs to runs/ (TensorBoard). Checkpoints save every 5k steps.

See ARCHITECTURE.md for hyperparameters and curriculum details.

---

Running the Benchmark

# Run the full 150-scenario evaluation
pixi run bash evaluation/run_final_eval.sh

# Results saved to results/final_evaluation.csv

---

Future Plans

→ Full roadmap

---

Contact

Sourav Kumar
📧 sourav710075@gmail.com
🐙 GitHub

Open to robotics engineering roles, autonomous systems research, and collaborations.

---

License

MIT License — see LICENSE for details.