Satellite Trajectory Projection and Collision Avoidance

Table of Contents

• Overview
• Features
• Environment Setup
• Installation
• Usage
• Data
• Performance Metrics
• Environment Design & Setup
• Reward Function Design
• Current Results
• Credits

Overview

This project implements an AI-powered satellite trajectory projection and collision avoidance system using reinforcement learning. It combines real-world satellite tracking data (TLE - Two-Line Element sets) with a Proximal Policy Optimization (PPO) model to predict and avoid potential collisions with space debris and other satellites.

Video Demonstration of Training

| Blue: Active Satellites in orbit | Red: Debris in orbit | Green: Predicted trajectory for model|

This animation demonstrates satellite trajectory adjustments with real-time collision avoidance.

Performance Metrics and Key Results

Evaluation summary (1,000 scenarios; deterministic seeds):

Policy	Success Rate	Collision Rate
PPO (learned)	~97.5%	~2.5%
Impulsive ΔV	~27.5%	~72.5%
Rule-based	~20.7%	~79.3%
No action	0%	100%

Notes:

• “Success” = survived to episode horizon without collision (some runs may end via fuel depletion).
• Artifacts include per-episode aggregates, time series, termination counts, and plots under results/evaluation/<timestamp>/.

Features

• Real-time TLE Data Fetching: Retrieves up-to-date satellite and debris data from CelesTrak.
• 3D Visualization: Provides an interactive 3D environment to visualize satellite orbits and debris.
• AI-powered Collision Avoidance: Utilizes a PPO model to dynamically avoid collisions.
• Dynamic Orbit Scaling: Adjusts orbit scales for improved visualization.
• Gravitational Influence Modeling: Accounts for Earth's and Moon's gravitational effects on satellite orbits.
• Performance Metrics Tracking: Logs and analyzes performance metrics during simulations.
• Deterministic Evaluator & Plots: Generates per-episode/per-policy CSVs and plots (collision rates, Δv usage, min distance, termination reasons).
• Per-Component Rewards: Exports survival, distance, closing, Δv penalties, smoothness, milestones, and collision penalties per episode.
• Multiple Baselines: Compare PPO against rule-based, impulsive ΔV, and no-action baselines.

Prerequisites

• Python 3.8+
• pip package manager

Installation

1. Clone the Repository

   git clone https://github.com/Logan-Cole-Luna/sat-trajectory-avoidance
   cd sat-trajectory-avoidance

2. Create Virtual Environment

   python -m venv venv
   source venv/bin/activate  # On Windows: venv\Scripts\activate

3. Install Dependencies

   ./setup.sh 

Usage

Download TLE Data from CelesTrack

python tle_export.py

Implement and Run the pretrained Satellite Model

python implement_sat.py

Train Your Own Satellite Model

python train_sat.py

Ealuate Satellite Model

python evaluate_sat.py --model models/satellite_avoidance_model.zip --episodes 1000

Outputs are written to results/evaluation/<timestamp>/ (ignored by git). You’ll get:

• policy_aggregates.csv, termination_reasons.csv, all_policies_per_episode.csv
• Optional time series and plots under plots/
• Reward component means and per-episode component CSVs

Data

The project uses TLE data from CelesTrak, stored in the tle_data/ directory. Ensure the TLE files are up-to-date for accurate simulations.

Performance Metrics

Performance metrics such as collision counts, reward accumulation, delta-v usage, and minimum distances to debris are logged and can be reviewed to evaluate the model's effectiveness.

Environment Design & Setup:

Reward Function Design

Current Results

Baselines Compared

• PPO (learned policy)
• Rule-based risk-aware controller (lateral push-away when within unsafe distance)
• Impulsive ΔV (finite-duration lateral thrust to approximate an instantaneous Δv pre-TCA)
• No action (do nothing)

Brief note on impulsive baseline: In classical astrodynamics, an “impulsive” burn models an instantaneous Δv with no position change; we approximate this with sustained lateral acceleration over several steps due to thrust limits.

Baseline Function For Evaluation:

• Velocity Adjustment: Adjusts velocity away from the nearest debris object if it is within a defined unsafe distance.
• No Action: Maintains current trajectory if no debris is within the unsafe distance.

Model Performance Overview

• PPO achieves ~97.5% collision-free episodes on 1,000 scenarios; rule-based ~20.7%; impulsive ΔV ~27.5%; no-action 0%.
• PPO yields larger average separation at closest approach and uses more Δv than heuristics (trading fuel for safety).
• Termination breakdowns and reward component tables are exported per evaluation run.

Error Analysis

• Delta-v Efficiency: While prioritizing collision avoidance, the PPO models consumed more delta-v than the baseline.
• Collision Detection Sensitivity: Needs refinement for better accuracy.
• Reward Function: Currently unbalanced and requires adjustments.
• Documentation: Additional documentation on fuel metrics is necessary.

Credits

• Data Sources: Credit to Celestrak for providing TLE data.