TreeSats 🛰️🌲

Defending Space Sovereignty

Satellites that protect themselves: end-to-end autonomous satellite collision avoidance in communication-denied environments

TreeSats enables satellites to detect collisions, execute evasive maneuvers, and assess threats—all without GPS or ground contact. Using only star tracker cameras (standard spacecraft hardware), TreeSats provides autonomous protection in contested space.

---

The Problem

Thousands of satellites orbit Earth today. SpaceX alone plans to deploy 1 million by 2030. Meanwhile, GPS and communications jamming is expanding across Eastern Europe, Southeast Asia, and Myanmar—affecting even commercial satellites. Ground station control isn't viable at this scale in contested regions.

Satellites are no longer passive—they maneuver unpredictably, creating collision risks with no margin for error.

The Solution:

TreeSats' Starguard system provides three autonomous capabilities:

1. Collision Detection - YOLOv8 + BoT-SORT identify satellites and debris in star tracker imagery, UKF estimates trajectories from angles-only measurements

2. Collision Avoidance - Convex optimization computes fuel-optimal evasive maneuvers for non-adversarial and adversarial satellites

3. Threat Assessment - NVIDIA Cosmos VLM conducts a satellite threat assessment for satellites in close proximity orbits

All processing happens onboard using existing star tracker cameras. No GPS. No ground link required.

---

Quick Start

# Install dependencies
pip install -r requirements.txt

# Generate synthetic training data (orbital sim + star tracker images)
python main.py --mode train --duration 0.5 --visualize

# Auto-label the generated images
cd sat/computer_vision
python auto_label.py batch data/images/train/ data/labels/train/

# Train YOLOv8 satellite detector
python train.py --epochs 150 --batch 64

# Run full pipeline (simulate → detect → track → avoid)
cd ../..
python main.py --mode pipeline

---

Key Features

• Communication-Independent: Operates without GPS or ground contact using only star tracker imagery
• Autonomous: Real-time collision detection, avoidance, and threat assessment without ground intervention
• Scalable: GPU acceleration handles 10,000+ satellites simultaneously
• Mission-Aware: Fuel-optimal maneuvers via convex optimization

---

Tech Stack

Component	Technology	Purpose
Orbital Propagation	Tensorgator (GPU)	High-fidelity Keplerian mechanics for 10k+ satellites
Object Detection	YOLOv8	Detect satellites/debris in star tracker images
Object Tracking	BoT-SORT	Persistent IDs across frames for trajectory extraction
State Estimation	Unscented Kalman Filter	Fuse measurements with orbital dynamics
Maneuver Planning	CVXPY	Convex optimization for collision avoidance
Classification	NVIDIA Cosmos VLM	Identify satellite types and assess threats

---

Repository Structure

treesats/
├── main.py                    # Entry point (train/pipeline modes)
├── sim/                       # GPU orbital simulation
│   ├── simulate.py           # Tensorgator propagation
│   └── star_tracker/         # Camera rendering
├── sat/
│   ├── computer_vision/      # YOLOv8 detection + tracking
│   │   ├── train.py
│   │   ├── auto_label.py
│   │   └── TRAINING.md       # Complete training guide
│   └── control/              # Collision avoidance + GNC
├── treesats_proximity/        # VLM threat assessment
└── flyby_data/               # Test imagery (Cassini, TIE, etc.)

---

Documentation

• Training Guide - YOLOv8 training for satellite detection
• Auto-Labeling Guide - Automated dataset generation
• Presentation Slides - Full system overview

---

Detailed Technical Overview

How Starguard Works

1. Detection Pipeline

Star trackers are attitude sensors present on virtually all spacecraft. TreeSats repurposes them for threat detection:

Star Tracker Image → YOLOv8 Detection → BoT-SORT Tracking → Bearing Angles

• Filters out background stars (known catalog)
• Detects moving objects as bright spots
• Assigns persistent track IDs
• Extracts pixel coordinates → elevation/azimuth angles

2. Angles-Only Navigation

From sequential star tracker images, TreeSats reconstructs full 6-DOF state:

Images	Information Extracted
1 photo	Bearing angles (el, az)
2 photos	Bearing + depth → Position vector
3 photos	Finite difference → Position + Velocity

Output: State vector x = [x, y, z, ẋ, ẏ, ż] in ECI frame

3. State Estimation (UKF)

An Unscented Kalman Filter fuses two information sources:

• Dynamics Model: Keplerian orbital propagation (M = E - e sin E)
• Measurements: Position/velocity from star tracker angles

This provides filtered, noise-resistant state estimates for collision prediction.

4. Collision Detection

Forward-propagate filtered trajectories to compute:

• Closest approach distance (miss distance)
• Time to closest approach
• Probability of collision (from covariance)

If risk exceeds threshold → trigger autonomous maneuver.

5. Evasive Maneuvers

Convex optimization (CVXPY) solves for minimum ΔV subject to:

• Collision avoidance constraint (minimum safe distance)
• Fuel budget limit
• Thrust magnitude/direction limits
• Mission constraints (maintain orbit parameters)

Output: Time-optimal thrust profile for safe evasion.

---

GPU Orbital Simulation

TreeSats uses Tensorgator (GPU-accelerated Keplerian propagator) to simulate 10,000+ satellites:

# Solve Kepler's equation: M = E - e sin E
# Distribute across 10k satellites on CUDA cores
positions, velocities = propagate_constellation(
    n_sats=10000,
    duration_hours=24,
    dt_seconds=10
)

Performance: 10,000 satellites × 8,640 timesteps (24h @ 10s) in seconds on GPU vs. hours on CPU.

---

Star Tracker Rendering

Pinhole camera model generates synthetic 256×256 images:

1. Project satellite positions into camera frame (RTN → body frame)
2. Apply perspective projection with field-of-view
3. Render as bright dots on black background
4. Save as training images

Auto-labeling: Brightness thresholding automatically generates YOLO bounding boxes from ground truth positions—no manual annotation required.

---

Computer Vision Pipeline

Training Workflow

# 1. Generate synthetic data
python main.py --mode train --duration 0.5 --visualize

# 2. Auto-label (brightness thresholding)
cd sat/computer_vision
python auto_label.py batch data/images/train/ data/labels/train/

# 3. Validate dataset
python prepare_data.py validate

# 4. Train YOLOv8
python train.py --epochs 150 --batch 64 --model yolov8m

Inference

from sat.computer_vision import detect_single, track

# Single image detection
result = detect_single("star_tracker_image.jpg")
print(f"Detected {result['count']} objects")

# Video tracking
results = track("star_tracker_sequence.mp4", tracker="botsort.yaml")

Output: Bounding boxes + persistent track IDs + pixel coordinates

---

VLM Threat Assessment

NVIDIA Cosmos Reason (served via vLLM) classifies satellites from imagery:

cd treesats_proximity
python assessment.py flyby_data/cassini/cassini.mp4

Identifies:

• Satellite type (Cassini spacecraft, communications sat, etc.)
• Object class (active satellite vs. debris vs. fairing)
• Threat level based on maneuver capability

vLLM provides high-throughput, low-latency inference for the vision-language model, enabling real-time proximity threat assessment onboard satellites.

Supports images, image sequences, and video (mp4/avi/mov).

---

Demonstration Scenarios

Chaser-Evader

• Chaser actively pursues target
• Evader (your satellite) detects and autonomously evades
• Real-time trajectory prediction enables proactive response

Multi-Object Tracking

• Simultaneous tracking of multiple threats
• Prioritized risk scoring (closest approach, time-to-collision)
• Coordinated maneuvers to avoid multiple objects

Collision Avoidance

• Forward propagate trajectories
• Predict collisions before they occur
• Compute fuel-optimal avoidance maneuver

---

Applications

Commercial Space

• Starlink, OneWeb, and mega-constellations
• Earth observation satellites
• Communications satellites

National Security

• Reconnaissance satellites in communication-denied regions
• Resilient operations in contested environments
• Autonomous collision detection, avoidance, and threat assessment

Space Situational Awareness

• Orbital debris tracking
• Traffic management for crowded orbits
• Collision warning systems

---

Team

• Sid Anantha - SpaceX Starship GNC
• Rahul Ayanampudi - Northrop Grumman GNC/Physical AI
• Lundeen Cahilly - Stanford CS, Incoming GNC @ K2 Space
• Raj Thapliyal - Defence Science Org. - Satellite Technology & Research / NUS ML

---

Roadmap

• Embedded deployment (spacecraft flight computer)
• Multi-sensor fusion (star tracker + IMU + magnetometer)
• Cooperative avoidance (inter-satellite coordination)
• Hardware-in-the-loop testing
• Flight software integration

---