MarsSimNav: Autonomous Terrain-Aware Rover Path Planning

MarsSimNav is a deep learning and path planning system designed to simulate autonomous Mars rover navigation using real NASA imagery. It leverages a DeepLabV3+ segmentation model trained on the AI4Mars dataset to classify Martian terrain types, followed by A search* to calculate optimal paths over hazardous terrain.

---

Features

• Semantic Segmentation: Powered by DeepLabV3+ with a ResNet50 backbone.
• Custom Pipeline: Specialized PyTorch Dataset handling for AI4Mars pixel-level annotations.
• Terrain-Aware Path Planning: Implements the A* algorithm via NetworkX.
• Cost-Based Navigation: Assigns navigation weights to different terrain types (soil, bedrock, sand, big rocks).
• Visualization: Generates visual overlays of the predicted rover path on original Mars surface images.
• Scalable: Supports subsetting for lightweight experimentation (e.g., 5,000 images).

---

Dataset

• Source: AI4Mars Dataset from NASA JPL.
• Imagery: Over 18,000 images captured by the Curiosity rover (MSL).
• Labels: Pixel-level annotations for:
  • Soil: Safe for traverse.
  • Bedrock: Stable but requires caution.
  • Sand: High risk of entrapment.
  • Big Rocks: Physical obstacles.

---

Technologies Used

• Core: Python 3.11, PyTorch, Torchvision
• Algorithms: DeepLabV3+, A* (NetworkX)
• Data & Visualization: Matplotlib, PIL, NumPy
• Environment: Google Colab / Linux

---

Getting Started

1. Clone the repository

git clone [https://github.com/pushkar-hue/marssimnav.git](https://github.com/pushkar-hue/marssimnav.git)
cd marssimnav

2. Install dependencies

pip install -r requirements.txt

3. Usage (Colab/Local)

• Ensure your trained model is available as deeplabv3_mars.pth.

• Place sample Mars images in ai4mars-subset/images/.

• Run the inference script or call the planning function:

Path Planning Logic

The system follows a 4-step pipeline to ensure safe navigation:

• Inference: Predict the segmentation mask using the DeepLabV3+ model.

• Mapping: Convert the mask into a Terrain Cost Map based on class IDs.

• Search: Use A* to find the lowest-cost path from start to goal coordinates.

• Overlay: Project the computed path onto the original image for visualization.

Terrain Cost

Terrain,Class ID,Cost Weight,Risk Level Soil,0,1,Low Bedrock,1,4,Moderate Sand,2,6,High Big Rock,3,10,Obstacle

Pixels with unknown labels (e.g. masked/NULL) are assigned high cost or skipped.

Sample Results

Input Image Segmentation Mask Planned Path Overlay mars_image.jpg predicted_mask.png planned_path_overlay.png

Future Work

• Add support for D* Lite and real-time replanning

• Interactive terrain selection and goal input via UI

• Animated simulation of rover motion along the path

• Deployable web dashboard using Streamlit or Gradio

• Integration with planetary data system (PDS) metadata

Authors

• Pushkar Sharma (@pushkar-hue)
• Ved Thorat (@i3hz)

Special thanks to NASA JPL for providing the AI4Mars Dataset.