Detour — On-Board AI Agents Saving Satellites from Orbital Debris

TreeHacks 2026 | NVIDIA Edge AI Track — Honourable Mention Winner

Devpost https://devpost.com/software/detour-64kpds?ref_content=user-portfolio&ref_feature=in_progress

Detour is an autonomous collision-avoidance system that runs on-board a satellite using NVIDIA's Nemotron LLM on the ASUS Ascent GX10 (Grace Blackwell). A multi-agent LangGraph pipeline detects debris threats, assesses risk, plans maneuvers, validates safety constraints, and executes avoidance burns — all locally with zero ground-station latency.

Architecture

┌──────────────────────────────────────────────────────────────────┐
│                    ASUS Ascent GX10 (On-Board)                   │
│                                                                  │
│  ┌─────────┐  ┌──────────┐  ┌──────────┐  ┌────────┐  ┌──────┐ │
│  │  SCOUT  │→ │ ANALYST  │→ │ PLANNER  │→ │ SAFETY │→ │ OPS  │ │
│  │ scan &  │  │ risk &   │  │ maneuver │  │ verify │  │BRIEF │ │
│  │ triage  │  │ refine   │  │ design   │  │& exec  │  │      │ │
│  └─────────┘  └──────────┘  └──────────┘  └────────┘  └──────┘ │
│       ↕             ↕             ↕             ↕               │
│  ┌──────────────────────────────────────────────────────────────┐   │
│  │              Physics Engine (deterministic)               │   │
│  │  screening · risk · CW dynamics · RK4 · SGP4 · Chan Pc   │   │
│  └──────────────────────────────────────────────────────────────┘   │
│       ↕                                                         │
│  ┌──────────────────────────────────────────────────────────────┐   │
│  │          Satellite Model (fuel, power, dynamics)          │   │
│  └──────────────────────────────────────────────────────────────┘   │
│       ↕                                                         │
│  ┌──────────────────────────────────────────────────────────────┐   │
│  │  Nemotron 3 Nano 30B (NVFP4) via vLLM — local inference  │   │
│  └──────────────────────────────────────────────────────────────┘   │
└──────────────────────────────────────────────────────────────────┘

Key Components

| Component | Path | Description | |-----------|------|-------------| | Agent Pipeline | agents/ | LangGraph 5-agent pipeline with tool-calling | | Physics Engine | engine/ | RK4 solver, J2 perturbation, CW dynamics, Chan collision probability | | Satellite Model | engine/models/active_satellite.py | Full orbital dynamics with resource management (fuel, power, battery) | | Tool Wrappers | agents/tools.py | 11 LangChain tools wrapping the physics engine | | API | api/ | FastAPI server with agent, catalog, conjunction, and satellite endpoints | | Frontend | frontend/ | Next.js + React Three Fiber 3D globe with live satellite tracking | | Ascent GX10 Setup | scripts/setup_gx10.sh | One-command setup for the ASUS Ascent GX10 |

Agent Pipeline

| Agent | Role | Tools | |-------|------|-------| | Scout | Scan catalog for upcoming conjunctions, triage by severity | scan_conjunctions, scan_demo_conjunctions | | Analyst | Deep risk assessment — Chan probability, high-fidelity TCA refinement | assess_risk, refine_conjunction, propagate_orbit | | Planner | Design avoidance maneuvers considering satellite resources | propose_avoidance_maneuvers, simulate_maneuver, get_satellite_status, check_maneuver_feasibility | | Safety | Validate constraints, approve or reject, execute approved burns | check_maneuver_constraints, get_satellite_status, check_maneuver_feasibility, execute_maneuver_on_satellite | | Ops Brief | Generate human-readable summary for operators | (synthesis only) |

Quick Start

1. Backend

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
uvicorn api.app:app --reload --port 8000

2. Frontend

cd frontend
npm install
npm run dev  # localhost:3000

3. Agent System (with Ascent GX10)

# Start Nemotron on the Ascent GX10
chmod +x scripts/setup_gx10.sh
./scripts/setup_gx10.sh

# Run agent pipeline
python -m agents.run "Scan for conjunction threats to satellite 25544 in the next 48 hours" --demo

4. Agent System (without GPU — dev mode)

# Set OPENAI fallback in .env
NEMOTRON_BASE_URL=https://api.openai.com/v1
NEMOTRON_API_KEY=sk-...
NEMOTRON_MODEL=gpt-4o-mini

python -m agents.run "Scan for threats" --demo

Model

nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4 — 4-bit quantized (NVFP4) for fast edge inference on the Ascent GX10. ~15GB model weight footprint, leaving ample memory for KV cache and concurrent requests on the 128GB unified memory Grace Blackwell SoC.

Served locally via NGC vLLM container with tool-calling (--enable-auto-tool-choice --tool-call-parser hermes --enable-chunked-prefill).

Why Edge AI?

| Ground Station | On-Board (Detour) | |---------------|-------------------| | 5-15 min communication delay | < 1 sec decision | | Limited pass windows | 24/7 monitoring | | Single point of failure | Autonomous operation | | Manual operator in the loop | Agent-validated decisions |

In LEO, a debris collision can happen in minutes. You can't wait for the next ground station pass.

Team

• Justyna — Frontend, 3D Visualization, UI/UX
• Ethan — ASUS Ascent GX10 Setup, Simulation Logic
• Adit — Satellite Data Feed, Simulation Logic
• Keanu — Ascent GX10 vLLM Setup, LangChain NVIDIA Nemotron Agent System