Natural language drone inspection powered by 8 NVIDIA AI technologies.
Speak a command. Simulate the mission. Inspect infrastructure. Generate reports.
See It In Action • The Problem • How It Works • NVIDIA Stack • Real-World Demo • Quick Start
Watch PARALLAX plan, simulate, and execute a real utility pole inspection in Seattle, WA — from a single English command to an AI-generated inspection report.
Unedited terminal recording of the complete 7-phase pipeline executing on an NVIDIA L40S GPU using Brev.
🖥️ Try the Interactive 3D Demo →
Interactive 3D visualization of the flight path planning. Uses simulated data for demonstration — no live API calls or GPU required. Clone the repo and open in your browser or click the hyperlink above.
I co-founded Voltair, an AI drone infrastructure startup, and hit a wall that every drone operator faces: the intelligence gap.
The hardware is incredible — modern drones fly autonomously for 40+ minutes with centimeter-level GPS accuracy. But commanding them? Still stuck in 2015:
- Operators spend hours manually programming waypoints
- Every flight requires a trained, FAA-certified pilot
- Footage is reviewed frame-by-frame by a human analyst
- One bad flight plan can crash a $10,000 drone in seconds
- If anything changes on site, you're starting from scratch
This doesn't scale. The US alone has 180 million utility poles, 600,000 bridges, and 5.5 million miles of power lines that need regular inspection. The bottleneck isn't the drone, it's the brain.
PARALLAX is the brain I wished we had at Voltair.
parallax "inspect all utility poles"
That's it. One command triggers a complete autonomous inspection pipeline:
The key differentiator: PARALLAX simulates every mission before the drone leaves the ground. Isaac Sim validates the entire flight path in a physics-accurate digital twin: checking collisions, altitude clearance, no-fly zones, and battery feasibility. If anything fails, the mission is rejected before you risk any hardware.
| Phase | Tool | What It Does |
|---|---|---|
| 0 | NeMoGuard | Validates the command is safe (rejects "crash the drone") |
| 1 | Nemotron via NIM | Parses natural language → structured mission with target assets |
| 2 | cuOpt | Optimizes visit order, minimizing flight distance (15% improvement) |
| 3 | Isaac Sim | Simulates full flight path in digital twin — collision & safety checks |
| 4 | Export | Generates GPS waypoints → Litchi CSV, DJI SDK, JSON |
| 5 | Drone | Executes validated mission autonomously |
| 6 | Cosmos Reason 2 | Analyzes captured footage with chain-of-thought visual reasoning |
| 7 | Report | Generates professional inspection report with severity ratings |
8 NVIDIA technologies working together on a single L40S GPU:
| Tool | Purpose | Integration |
|---|---|---|
| NeMoGuard | Safety & content filtering | Real API — validates every command |
| Nemotron (mini-4b-instruct) | Natural language understanding | Real API via NIM |
| NIM | Model inference microservices | Powers Nemotron inference |
| cuOpt | Route optimization | Nearest-neighbor with haversine distance |
| Isaac Sim | Physics simulation & validation | Real on Brev — flight path validation |
| Omniverse | Digital twin platform | Via Isaac Sim environment |
| Cosmos Reason 2 (8B) | Physical AI video reasoning | Real on Brev — drone footage analysis |
| Brev | GPU cloud infrastructure | NVIDIA L40S, 48GB VRAM |
Model weights: nvidia/Cosmos-Reason2-8B on Hugging Face
PARALLAX was tested on real infrastructure in Seattle, WA.
Four utility poles were pre-mapped with GPS coordinates. The pipeline generated a validated 19-waypoint flight plan. A DJI Mini 2 captured inspection footage which was analyzed by Cosmos Reason 2 running on an NVIDIA L40S GPU.
Results:
- Cosmos identified all 4 poles individually using chain-of-thought reasoning
- Flagged vegetation encroachment warnings on Poles 1 and 3
- Cleared Poles 2 and 4 with no issues
- Generated a complete inspection report with severity ratings and recommended actions
GPS waypoints exported to Litchi Mission Hub — real satellite imagery of the inspection site
# Clone
git clone https://github.com/HMAC10/Parallax.git
cd Parallax
# Run with mock data (no GPU needed)
python parallax_pipeline.py --command "inspect all poles" --site demo_site.json --mock
# Run with real Cosmos Reason 2 (requires 32GB+ VRAM)
python parallax_pipeline.py --command "inspect all poles" \
--site demo_site.json \
--footage drone_video.mp4
# Simple shortcut
parallax "inspect all utility poles"
# Test safety rejection
python parallax_pipeline.py --command "crash the drone" --site demo_site.json- Python 3.10+
- NVIDIA GPU with 32GB+ VRAM (for Cosmos Reason 2)
pip install -r requirements_cosmos.txt
Mock mode (--mock) runs the full pipeline without GPU requirements.
parallax_pipeline.py Main end-to-end pipeline (7 phases)
cosmos_analyzer.py Cosmos Reason 2 video/image analysis
report_generator.py Inspection report generator (Markdown + JSON)
site_config.py Site configuration with GPS asset mapping
export_waypoints.py Waypoint exporter (Litchi CSV, JSON, DJI SDK)
demo_site.json Real site config: 4 utility poles in Seattle
web_demo.html Interactive 3D flight path visualization (Three.js)
- Real-time video streaming — Cosmos analyzes footage during the flight, not after, giving engineers a live AI-assisted view of what the drone sees
- Automated airspace checks — B4UFLY integration directly into the pipeline so PARALLAX automatically verifies restricted airspace, TFRs, and no-fly zones before launch — no pilot involvement needed
- Dynamic path adjustment — real-time obstacle avoidance where the drone creates new paths on the fly if something unexpected enters its flight path, rather than blindly following pre-programmed waypoints
- Professional report generation — exportable PDFs ready for engineering teams, insurance adjusters, and regulatory compliance
- Automated 3D environment generation — using satellite imagery and uploaded photos to automatically build the digital twin, so any new site can be validated without manual 3D modeling
- Multi-drone fleet coordination — multiple drones sharing detections, handing off monitoring zones, and collaborating on large-scale inspections from a single command
Today, drone inspection requires a trained pilot, a mission planner, and an analyst. PARALLAX collapses all three roles into a single natural language command.
The long-term vision: any utility worker, construction manager, or insurance adjuster can inspect infrastructure by simply telling the drone what to look at. One operator managing a fleet of autonomous inspection drones — that's the future PARALLAX is building toward.
Built by Hunter McKay - University of Washington Double Major in Information Systems & Finance (3.98 GPA). Co-founded Voltair, an AI drone infrastructure startup that won the $25K Dempsey Startup Competition (1st of 174 teams) and $15K Environmental Innovation Challenge.
PARALLAX was born from a real problem encountered building Voltair - the drone industry has incredible hardware but lacks the AI brain to make autonomous inspection accessible to everyone.