World Models & AI for Autonomous Vehicles: Airside Reference ODD Master Synthesis

Date: 2026-03-21 Scope: Comprehensive research survey across 20 deep-dive reports covering world models, VLAs, simulation, end-to-end driving, and their applicability to autonomous vehicles. Airport airside operations are the detailed reference ODD in this synthesis, not the default lens for the whole repository.

---

Table of Contents

1. Executive Summary
2. The Landscape: What Exists Today
3. Top Candidate Approaches for Airside AV
4. NVIDIA Alpamayo: Deep Dive
5. Recommended Architecture
6. The Airside Data Problem & Solutions
7. Simulation Strategy
8. Safety & Certification Path
9. Hardware & Deployment
10. Research Report Index
11. Key Papers & Repos
12. Strategic Recommendations

---

1. Executive Summary

Read this page as an airside reference case for a generic AV stack. The same stack questions should be re-evaluated for road AVs, warehouses, logistics yards, ports, mines, construction sites, farms, delivery robots, and campuses before deployment claims are transferred.

The autonomous driving field has undergone a paradigm shift from modular perception-planning-control pipelines toward learned world models and end-to-end architectures. Three converging trends make this the right time to adopt these approaches for airside AV:

1. World models are production-ready. Comma.ai's openpilot v0.11 ships a 2B-parameter diffusion transformer world model — the first robotics agent trained fully in learned simulation. Wayve's GAIA series (up to 15B params) navigates 500+ cities with a single model.

2. NVIDIA Alpamayo provides a research foundation. Released January 2026, this 10B-parameter VLA (8.2B Cosmos-Reason backbone + 2.3B action expert) with Chain-of-Causation reasoning, 1,727 hours of driving data, and AlpaSim closed-loop simulator. Important: Model weights are non-commercial (research/eval only), and it is camera-only — designed as a teacher model for distillation to smaller edge-deployable models.

3. Airside is an ideal ODD for world models. Low speeds (5-30 km/h), structured environments, and the absence of large-scale airside datasets (a weakness for traditional approaches) becomes a strength for world models that can leverage pre-training on road driving and adapt with minimal domain-specific data.

The Core Thesis

A world model pre-trained on road driving data, fine-tuned on airside data, and combined with occupancy-based planning can leapfrog traditional airside AV approaches that rely on hand-crafted rules and HD maps.

Key advantages over current airside AV solutions (TractEasy, reference airside AV stack, etc.):

• No per-class object detection needed — occupancy prediction is class-agnostic, handling aircraft, GSE, personnel without per-type training
• Generalization across airports — world models + VLAs can adapt to new layouts with minimal data, vs. re-mapping for each airport
• Language-grounded reasoning — VLAs can process ground control instructions and provide explainable decisions for regulatory compliance
• Synthetic data generation — world models generate training scenarios, addressing the zero-public-dataset problem for airside

---

2. The Landscape

2.1 World Model Architectures (Maturity Ranking)

Approach	Maturity	Key Systems	Domain Fit Notes
Latent world models + RL	Production	DreamerV3/V4, TD-MPC2, Think2Drive	High — proven for control
VLA (Vision-Language-Action)	Near-production	Alpamayo, LINGO-2, DriveVLM, pi0	Very High — language grounding
Diffusion world models	Research+	GAIA-2, Vista, DriveDreamer-2, DIAMOND	High — simulation & planning
Occupancy world models	Research+	OccWorld, Drive-OccWorld, OccSora	Very High — class-agnostic
Autoregressive video models	Research	DrivingGPT, DrivingWorld, Copilot4D	Medium — simulation
JEPA (embedding prediction)	Early research	V-JEPA 2, AD-L-JEPA	High potential — efficiency
3DGS/NeRF reconstruction	Research+	Street Gaussians, SplatAD, PVG	High — airport digital twins

2.2 Current Airside AV Players

Company	Product	Autonomy	Deployments	Tech Stack
TractEasy (TLD + EasyMile)	EZTow, EZDolly	L4	Narita, Changi, Munich, Dubai	GPS + 3D LiDAR + fusion
reference airside AV stack	autonomous baggage/cargo tug, autonomous cargo vehicle	L4	Zurich, Schiphol, Heathrow	LiDAR + 360 cam + GPS + IMU
Charlatte Autonom	AT135	L4	CDG (Air France), Frankfurt	V2X + sensor fusion
Fernride	Teleoperation + autonomy	L4 tele	NVIDIA partnership	Progressive autonomy
Ohmio	Autonomous shuttles	L4	JFK, Schiphol, Brussels	Undisclosed

Critical gaps in current solutions:

• All use traditional perception pipelines — no world models
• No public airside driving datasets exist
• FAA CertAlert 24-02 (Feb 2024) is the only formal US guidance
• No A-SMGCS integration with autonomous GSE
• Limited adverse weather handling (de-icing, jet blast, FOD)

2.3 Regulatory Landscape

Standard	Status	Relevance
FAA CertAlert 24-02	Only formal US guidance (2024)	Acknowledges AGVS, standards in development
ISO 3691-4:2020	Current certification path	Driverless industrial trucks — used by TractEasy/reference airside AV stack
ISO/PAS 8800	Published Dec 2024	AI safety lifecycle for AV — bridges ISO 26262 gap
EASA AI Roadmap 2.0	In progress, targeting 2028	W-shaped development process for aviation AI
ISO 21448 (SOTIF)	Active	Safety of intended functionality — key for ML models
UL 4600	Active	Safety case-based evaluation of autonomous products

---

3. Top Candidate Approaches for Airside AV

3.1 Tier 1: Build On Now

A. NVIDIA Alpamayo + AlpaSim (Recommended Starting Point)

• What: 10.5B VLA (8.2B Cosmos-Reason backbone + 2.3B action expert)
• Why: Open-source, Chain-of-Causation reasoning generates trajectories AND explanations, 1,727h driving data across 25 countries, AlpaSim for closed-loop testing
• Alpamayo 1.5: Adds RL post-training, flexible multi-camera support, text-guided planning
• Airside path: Fine-tune on airside data, leverage language reasoning for ground instructions, use AlpaSim for scenario testing
• License: Apache 2.0 + NVIDIA Open Model License
• Partners: Lucid, JLR, Uber, Berkeley DeepDrive

B. Occupancy World Models (OccWorld / Drive-OccWorld)

• What: Predict future 3D occupancy grids — class-agnostic scene prediction
• Why: No per-class detection needed. Handles aircraft, GSE, personnel, FOD without type-specific training. Drive-OccWorld (AAAI 2025) adds action conditioning — 33% improvement over UniAD
• Airside path: Pre-train on nuScenes/Waymo occupancy, fine-tune on airside. Jet blast zones become hazard occupancy predictions
• Open source: OccWorld on GitHub (ECCV 2024)

C. NVIDIA Cosmos World Foundation Models

• What: Open-weight world foundation models (4-14B params, trained on 9000T+ tokens)
• Why: Pre-trained world models for fine-tuning on domain-specific data. Up to 2048x compression tokenizers
• Airside path: Fine-tune Cosmos on airside video for simulation and data augmentation
• License: Apache 2.0 + NVIDIA Open Model

3.2 Tier 2: Integrate Short-Term

D. 3D Gaussian Splatting for Airport Digital Twins

• What: Reconstruct airport environments from sensor logs for photorealistic simulation
• Key systems: Street Gaussians (135 FPS, 30min training), SplatAD (CVPR 2025, joint camera+LiDAR)
• Airside path: Drive around airport, reconstruct as 3DGS scene, use for closed-loop testing with counterfactual scenarios (insert aircraft, simulate pushback, test FOD response)

E. Foundation Model Perception (Open-Vocabulary Detection)

• What: SAM, Grounding DINO, YOLO-World for detecting objects never seen in training
• Why: Airside has unique object classes (30+ types of GSE, 100+ aircraft variants). Open-vocabulary detection handles these without per-class annotation
• Airside path: Zero-shot detection with text prompts ("baggage tractor", "jet bridge", "belt loader"), then fine-tune with LoRA on small airside dataset
• Data estimate: 0-50 examples per class for zero-shot; 200-500 for fine-tuned 80%+ mAP

F. Map-Free Driving (MapTR/MapTRv2)

• What: Online vectorized map construction from sensors — no HD map needed
• Why: Airport layouts change frequently (construction, seasonal operations, NOTAMs). HD maps are expensive to maintain per-airport
• Airside path: Combine with AIXM/AMXM airport data as prior, NOTAM integration for dynamic restrictions

3.3 Tier 3: Research & Future

G. JEPA-Style World Models (V-JEPA 2, AD-L-JEPA)

• What: Predict future scene embeddings, not pixels — ~15x faster planning than video-generation baselines
• Why: Massive efficiency gains. V-JEPA 2-AC plans in 16 seconds vs. 4 minutes for pixel-generation baselines
• Status: Early but Meta open-sourcing aggressively. AD-L-JEPA is first driving-specific JEPA (AAAI 2026)

H. LLM-Based Planning & ATC Integration

• What: LLMs for reasoning about traffic rules, ground control instructions, edge cases
• Key: DriveReg achieves 100% traffic rule compliance via RAG. Delft's LLM-based ATC agent resolved 119/120 conflicts
• Airside path: RAG over ICAO/FAA rules + airport-specific procedures. Process digital taxi instructions (NASA NLU work exists)

I. Multi-Agent Fleet Coordination

• What: Shared world models across vehicle fleet, cooperative perception, turnaround sequencing
• Key: Moonware HALO (deployed at US hubs, 20% delay reduction), SuperMap Apron Commander
• Airside path: Fleet world model shared via 5G/CBRS, A-CDM integration for turnaround optimization

---

4. NVIDIA Alpamayo: Deep Dive

Alpamayo is the most directly applicable system for airside AV development. Here's why:

Architecture

Input: Multi-camera images + vehicle state
    |
    v
[Cosmos-Reason Backbone (8.2B)] -- Vision encoder + language model
    |
    v
[Chain-of-Causation Reasoning] -- Generates causal explanation of scene
    |                              ("Vehicle ahead is braking because...")
    v
[Action Expert (2.3B)] -- Generates driving trajectories
    |
    v
Output: Trajectory + Natural language explanation

Key Capabilities

• Chain-of-Causation reasoning: Generates both trajectories AND interpretable explanations — critical for regulatory compliance in aviation
• Text-guided planning (v1.5): Can follow natural language instructions — maps directly to ground control instructions
• RL post-training (v1.5): Reinforcement learning fine-tuning improves real-world performance
• 700K reasoning traces: Pre-built dataset for reasoning-aware training
• AlpaSim: Microservice-based closed-loop simulator for testing

Airside Adaptation Strategy

1. Phase 1 (0-3 months): Deploy Alpamayo in shadow mode on airside vehicles. Collect data with explanations.
2. Phase 2 (3-6 months): Fine-tune on collected airside data. Add airport-specific reasoning traces (jet blast awareness, FOD detection, stand procedures).
3. Phase 3 (6-12 months): RL post-training on airside scenarios via AlpaSim. Add ground control instruction following.
4. Phase 4 (12+ months): Multi-airport deployment with few-shot adaptation.

---

5. Recommended Architecture

5.1 Layered Architecture for Airside AV

┌─────────────────────────────────────────────────┐
│              SAFETY LAYER (Always Active)         │
│  Simplex architecture: High-performance + safe    │
│  RSS-based safety envelope │ OOD detection        │
│  Teleoperation fallback (Fernride-style)          │
└───────────────────────┬─────────────────────────┘
                        │
┌───────────────────────▼─────────────────────────┐
│           PLANNING & REASONING LAYER              │
│  Alpamayo VLA (trajectory + explanation)           │
│  OR Occupancy-based MPC (Drive-OccWorld)           │
│  LLM reasoning for edge cases (RAG over rules)    │
│  Multi-agent coordination (fleet world model)     │
└───────────────────────┬─────────────────────────┘
                        │
┌───────────────────────▼─────────────────────────┐
│           WORLD MODEL LAYER                       │
│  4D Occupancy prediction (OccWorld-style)          │
│  Motion prediction (MotionLM / MTR++)              │
│  Scene understanding (DINOv2 + open-vocab det.)   │
│  Online mapping (MapTRv2 + AIXM prior)            │
└───────────────────────┬─────────────────────────┘
                        │
┌───────────────────────▼─────────────────────────┐
│           PERCEPTION LAYER                        │
│  BEV fusion (BEVFormer/BEVFusion)                 │
│  Multi-sensor: cameras + LiDAR + 4D radar         │
│  Foundation model features (DINOv2 backbone)      │
│  Open-vocabulary detection (Grounding DINO)       │
└───────────────────────┬─────────────────────────┘
                        │
┌───────────────────────▼─────────────────────────┐
│           SENSOR & LOCALIZATION LAYER             │
│  Cameras (surround view) + LiDAR + 4D radar       │
│  RTK-GNSS + visual/LiDAR SLAM fallback           │
│  UWB beacons for GPS-degraded areas              │
│  ADS-B receiver for aircraft awareness            │
└─────────────────────────────────────────────────┘

5.2 Integration Points

Airport System	Integration Method	Purpose
A-SMGCS	API / ASTERIX Cat 010/062	Aircraft position awareness
A-CDM / AODB	ACRIS Semantic Model	Flight schedule, turnaround timing
NOTAM system	AIXM/AMXM parser	Dynamic restriction awareness
Digital tower	Data feed	Surface movement clearances
ADS-B	1090ES receiver	Aircraft proximity detection
Fleet management	5G/CBRS mesh	Vehicle coordination

5.3 Sensor Suite Recommendation

Sensor	Qty	Purpose	Airside Rationale
Cameras (surround)	6-8	Primary perception	Rich semantic info, low cost
LiDAR (128-ch)	1	3D geometry, localization	Handles reflective surfaces, large objects
4D Radar	2-4	Velocity, all-weather	Works in rain, de-icing spray, fog
RTK-GNSS	1	Primary localization	cm-level in open areas
UWB anchors	Site-based	GPS fallback	Near terminals, under aircraft
ADS-B receiver	1	Aircraft awareness	Real-time aircraft positions
IMU	1	Dead reckoning	Bridge GPS gaps

---

6. The Airside Data Problem & Solutions

6.1 The Problem

Zero public airside driving datasets exist. The only resources:

• Synth_Airport_Taxii (synthetic, limited)
• AssistTaxi (marking recognition, limited)
• AeroVect proprietary (not available)

6.2 The Solution: 4-Phase Data Engine

Phase	Timeline	Strategy	Data Volume
1. Foundation	0-6 months	Pre-train on road driving (nuScenes, Waymo, nuPlan). Transfer learn with domain adaptation	10,000+ hours (existing)
2. Bootstrap	3-9 months	Deploy in shadow mode at 1-2 airports. Manual annotation + auto-labeling (SAM + Grounding DINO)	100-500 hours
3. World Model Training	6-18 months	Train world model on airside data. Generate synthetic scenarios via Cosmos/GAIA. 3DGS digital twin of airports	1,000+ hours (real + synthetic)
4. Scaling	12+ months	Active learning from fleet. Federated learning across airports. Continuous world model improvement	5,000+ hours

6.3 Synthetic Data Generation Pipeline

Real Airport Scan (3DGS/NeRF)
    → Reconstructed Digital Twin
    → Insert synthetic agents (aircraft, GSE, personnel)
    → Vary weather, lighting, time-of-day
    → Render multi-sensor data (camera + LiDAR + radar)
    → World model training data

Tools: NVIDIA Cosmos for video generation, 3DGS for scene reconstruction, DriveDreamer-2 for LLM-prompted scenario generation, Unreal Engine with airport assets for structured simulation.

6.4 Domain Adaptation Strategy

Pre-training curriculum:

1. General vision (DINOv2 on ImageNet)
2. Driving video (Cosmos on road data)
3. Structured low-speed driving (parking lots, industrial areas)
4. Synthetic airport scenes
5. Real airport data (fine-tuning)
6. Continual learning from deployment

---

7. Simulation Strategy

7.1 Multi-Fidelity Simulation Stack

Level	Tool	Fidelity	Speed	Purpose
L1: Unit test	AlpaSim	Low	Real-time	Component testing
L2: Scenario	CARLA + airport assets	Medium	10-100x	Scenario validation
L3: Neural sim	Cosmos / GAIA	High visual	1-10x	Perception testing
L4: Digital twin	3DGS reconstruction	Photorealistic	Real-time	Closed-loop testing
L5: Shadow mode	Real vehicle	Real	1x	Pre-deployment validation

7.2 Key Scenarios to Simulate

Scenario	Priority	Challenge
Aircraft pushback (nose-in/nose-out)	Critical	Large dynamic object, jet blast
Crossing active taxiway	Critical	ATC clearance required
FOD on apron	High	Small object detection at distance
De-icing operations	High	Glycol spray, reduced visibility
Multi-vehicle turnaround sequencing	High	Coordination, timing
Emergency vehicle response	Critical	Yield behavior
Night operations with apron lighting	Medium	Lighting variation
Snow/ice operations	Medium	Traction, visibility
Personnel walking near aircraft	Critical	Vulnerable road user prediction

---

8. Safety & Certification Path

8.1 Recommended Certification Framework

Given that airside AV falls between automotive and aviation domains:

1. Primary: ISO 3691-4:2020 (driverless industrial trucks) — used by current deployments
2. AI safety: ISO/PAS 8800 (AI safety lifecycle) — newly published Dec 2024
3. Functional safety: ISO 26262 adapted for airside ODD
4. Intended functionality: ISO 21448 (SOTIF) — critical for ML-based perception
5. Safety case: UL 4600 framework with GSN notation
6. Future-proofing: Track EASA AI Roadmap 2.0 for aviation-specific AI certification

8.2 Safety Architecture for World Models

┌──────────────────────────────────────┐
│         WORLD MODEL PLANNER          │
│  (High performance, learned)         │
├──────────────────────────────────────┤
│         SAFETY MONITOR               │
│  - OOD detection (ensemble)          │
│  - Confidence calibration            │
│  - RSS safety envelope check         │
│  - Occupancy prediction validation   │
├──────────────────────────────────────┤
│      DECISION: Safe? ─── Yes ──→ Execute world model trajectory
│                  │
│                  No
│                  ↓
│         SAFE FALLBACK CONTROLLER     │
│  (Rule-based, formally verified)     │
│  - Graceful stop                     │
│  - Maintain safe distance            │
│  - Request teleoperation             │
└──────────────────────────────────────┘

8.3 AMLAS (Assurance of ML for Autonomous Systems) Process

1. ML Safety Requirements → derive from SOTIF analysis
2. ML Data Management → curate airside dataset with bias analysis
3. ML Model Learning → train with safety-aware objectives (SafeDreamer)
4. ML Model Verification → test against safety requirements
5. ML Model Deployment → runtime monitoring + OOD detection
6. ML Model Operation → continuous validation from fleet data

---

9. Hardware & Deployment

9.1 Compute Platform Recommendation

Option	TOPS	Power	Cost	Recommendation
NVIDIA Orin	275	60W	~$1,500	Best for initial deployment — proven, Alpamayo compatible
NVIDIA Thor	2,000+	130W	TBD	Future upgrade — runs full world models on-vehicle
Qualcomm Ride	1,280 (dual)	Varies	Competitive	Alternative if NVIDIA supply constrained

NVIDIA Orin (60W) is negligible on electric GSE with 30-100 kWh batteries (< 0.2% of battery capacity per hour).

9.2 Cloud-Edge Hybrid Architecture

Workload	Location	Latency	Rationale
Perception + planning	On-vehicle (Orin)	<100ms	Safety-critical, must be real-time
World model inference	On-vehicle (Orin)	<200ms	Latent-space models fit in 8GB
World model training	Cloud/edge server	Async	Too compute-intensive for on-vehicle
Simulation (AlpaSim)	Airport edge server	N/A	Scenario testing, not real-time
Fleet coordination	Airport edge server	<50ms	5G/CBRS connectivity
Data upload & labeling	Cloud	Async	Large data volumes (1-4 TB/hour)

9.3 Fleet Cost Model (50 vehicles)

Component	Year 1 Cost	Notes
Compute (Orin per vehicle)	$75K	Hardware + integration
Sensors per vehicle	$30-50K	LiDAR + cameras + radar + GNSS
5G/CBRS infrastructure	$200-400K	Airport-wide coverage
Edge server	$100-200K	Training + fleet coordination
Cloud (training)	$200-500K	GPU cluster rental
Software development	$1-5M	World model adaptation, integration
Total Year 1	$2.1-9.2M
Payback	1-3 years	Via labor cost savings

---

10. Research Report Index

Core World Models & Architecture

Report	File	Key Findings
World Models for AV	30-autonomy-stack/world-models/overview.md	17+ architectures, readiness assessment
End-to-End AV	30-autonomy-stack/end-to-end-driving/e2e-architectures.md	SparseDrive SOTA (0.06% collision), recommended architecture
Diffusion World Models	30-autonomy-stack/world-models/diffusion-world-models.md	Sora, Genie 2/3, 13 driving-specific models
RL with World Models	30-autonomy-stack/world-models/rl-with-world-models.md	Dreamer v1-v4, TD-MPC2, SafeDreamer
Tokenized & JEPA	30-autonomy-stack/world-models/tokenized-and-jepa.md	VQ-VAE world models, JEPA ~15x faster planning
4D Occupancy	30-autonomy-stack/world-models/occupancy-world-models.md	Class-agnostic prediction, jet blast as hazard occupancy
Occupancy Comparison	30-autonomy-stack/world-models/occupancy-networks-comparison.md	20 methods compared, FlashOcc 197.6 FPS
Open-Source Repos	30-autonomy-stack/world-models/opensource-implementations.md	21 repos rated, only 6 fully usable

Models & Approaches

Report	File	Key Findings
VLAs	30-autonomy-stack/vla-vlm/vla-for-driving.md	15+ VLA models, LINGO-2, DriveVLM
Alpamayo	30-autonomy-stack/vla-vlm/alpamayo-setup.md	Camera-only, non-commercial, teacher model for distillation
Company Deep Dives	30-autonomy-stack/end-to-end-driving/company-approaches.md	Wayve/NVIDIA/Tesla/Waymo/Comma.ai strategies
LLM Reasoning	30-autonomy-stack/planning/llm-reasoning-planning.md	DriveReg 100% rule compliance, Moonware HALO
Motion Prediction	30-autonomy-stack/planning/motion-prediction.md	MotionLM, gate turnaround sequencing
Foundation Models	30-autonomy-stack/perception/overview/vision-foundation-models.md	SAM, DINO, open-vocab detection
DINOv2	30-autonomy-stack/perception/overview/dinov2-foundation-models-driving.md	Adapter-mediated integration, LoRA rank 32 optimal

Simulation & Data

Report	File	Key Findings
Neural Simulation	30-autonomy-stack/simulation/neural-simulation-platforms.md	Alpamayo, Cosmos, 12+ sim platforms
3DGS & NeRF	30-autonomy-stack/simulation/neural-scene-reconstruction.md	Street Gaussians 135 FPS, SplatAD
Airport Digital Twins	30-autonomy-stack/simulation/airport-digital-twins.md	Autonoma, SITA, $50K-$4M
Simulators for Airside	30-autonomy-stack/simulation/simulators-for-airside.md	CARLA/AWSIM/Isaac Sim comparison
Datasets & Benchmarks	50-cloud-fleet/data-platform/data-engines-datasets.md	Zero airside datasets, 4-phase bootstrap plan
Synthetic Data	50-cloud-fleet/data-platform/synthetic-data-generation.md	Cosmos +16.2% mAP foggy, 7-phase pipeline

Deployment & Safety

Report	File	Key Findings
Safety & Certification	60-safety-validation/standards-certification/certification-guide.md	ISO/PAS 8800, EASA AI Roadmap, AMLAS
ISO 3691-4	60-safety-validation/standards-certification/iso-3691-4-deep-dive.md	$130K-380K, harmonized May 2024, 27 safety functions
Regulatory Trajectory	80-industry-intel/regulations/regulatory-trajectory-deep-dive.md	FAA AC ~2028-2029, EASA ~2028
Ground Crew Safety	70-operations-domains/airside/safety/ground-crew-pedestrian-safety.md	27K accidents/yr, hi-vis paradox
Insurance & Liability	80-industry-intel/regulations/insurance-liability-airside.md	$35M per engine exposure
Compute & Hardware	20-av-platform/compute/edge-platforms.md	Orin 275 TOPS @ 60W, fleet cost model
NVIDIA Orin	20-av-platform/compute/nvidia-orin-technical.md	8 power modes, DLA 74% at 15W
NVIDIA Thor	20-av-platform/compute/nvidia-drive-thor.md	~1000 TOPS, FP8 native, 2025+
TensorRT Guide	20-av-platform/compute/tensorrt-deployment-guide.md	PointPillars 6.84ms, DLA deployment
4D Radar	20-av-platform/sensors/4d-radar.md	Continental ARS548, immune to all weather
Airside Industry	70-operations-domains/airside/operations/industry-overview.md	21 companies, FAA CertAlert 24-02
Multi-Agent & Fleet	30-autonomy-stack/multi-agent-v2x/fleet-coordination.md	V2X, Moonware, 5G/CBRS, A-CDM
Mapping & Localization	30-autonomy-stack/localization-mapping/overview/mapping-and-localization.md	Map-free driving, AIXM, NOTAM integration
Robustness	60-safety-validation/verification-validation/robustness/adverse-conditions.md	De-icing, jet blast, FOD, 4D radar

---

11. Key Papers & Repos

Must-Read Papers

1. Alpamayo — NVIDIA's 10.5B VLA for L4 AV (CES 2026)
2. GAIA-1/2/3 — Wayve's world model series (9B-15B params)
3. DreamerV3/V4 — Model-based RL at scale (Nature 2025)
4. OccWorld — 3D occupancy world model (ECCV 2024)
5. Drive-OccWorld — Action-conditioned occupancy (AAAI 2025)
6. EMMA — Waymo's end-to-end multimodal model (on Gemini)
7. Cosmos — NVIDIA world foundation models (9000T+ tokens)
8. SafeDreamer — Safe RL with world models (ICLR 2024)
9. WorldRFT — RL fine-tuning for world models (AAAI 2026, 83% collision reduction)
10. V-JEPA 2 — Meta's efficient video world model (2025)

Key Open-Source Repos

Repo	Description
NVIDIA Alpamayo	10.5B VLA + AlpaSim
NVIDIA Cosmos	World foundation models
OpenDWM	Driving world model framework
OccWorld	3D occupancy world model
DriveDreamer	World model from real driving
DiffusionDrive	Real-time E2E driving (CVPR 2025)
DIAMOND	Diffusion RL world model
V-JEPA 2	Efficient video prediction
AD-L-JEPA	JEPA for driving LiDAR
openpilot	Open-source E2E driving
Awesome-World-Model	Paper collection

---

12. Strategic Recommendations

Immediate Actions (0-3 months)

1. Set up Alpamayo + AlpaSim — Download, run inference, understand the architecture
2. Begin airside data collection — Mount cameras on existing GSE in shadow mode
3. Build 3DGS digital twin — Scan your primary airport for simulation
4. Deploy open-vocab detection — Grounding DINO for zero-shot airside object detection

Short-Term (3-12 months)

5. Fine-tune Alpamayo on airside data — Start with simple scenarios (straight-line baggage tug routes)
6. Train occupancy world model — Pre-train on nuScenes, fine-tune on airside
7. Build synthetic data pipeline — Cosmos for video generation, 3DGS for scenarios
8. Engage with FAA/EASA — Begin safety case development using AMLAS methodology

Medium-Term (12-24 months)

9. RL post-training — Fine-tune world model policy with SafeDreamer-style constraints
10. Multi-airport adaptation — Test few-shot generalization to new airports
11. Fleet coordination — Shared world model via 5G, A-CDM integration
12. Ground control instruction following — LLM-based digital taxi instruction processing

Long-Term (24+ months)

13. Full autonomous turnaround sequencing — Multi-agent world model for coordinated GSE
14. Continuous learning data engine — Fleet-wide active learning, federated model updates
15. Certification — ISO 3691-4 + ISO/PAS 8800 + SOTIF safety case

Key Risk Mitigations

Risk	Mitigation
No airside training data	Synthetic data + transfer learning + active learning
Regulatory uncertainty	Build safety case with established frameworks (AMLAS, UL 4600)
World model hallucination	Simplex architecture with verified fallback controller
Compute constraints	Start with Orin (latent models), upgrade to Thor for full models
Airport stakeholder buy-in	Demonstrate in shadow mode first, show explainable decisions via VLA

---

This synthesis draws from 159 research documents totaling ~134,000 lines of analysis across 300+ papers, models, and systems. See INDEX.md for topic-based navigation.

Related Documents

Next Step	Document
Start coding today	90-synthesis/master/getting-started.md
Choose POCs	90-synthesis/poc-roadmaps/poc-proposals.md
Assess readiness	90-synthesis/readiness-risk/technology-readiness.md
Understand competition	80-industry-intel/market-competitive/competitive-landscape.md
Manage risks	90-synthesis/readiness-risk/risk-register.md
Full architecture	90-synthesis/decisions/design-spec.md