Back to Blog
April 29, 2026 • Tim Fernholz

Scout AI Raises $100M to Train War-Ready Autonomous Vehicle Models

Scout AI Raises $100M to Train War-Ready Autonomous Vehicle Models Scout AI Raises $100M to Train War-Ready Autonomous Vehicle Models Scout AI Raises $100M to Train War-Ready Autonomous Vehicle Models

Scout AI Raises $100M to Train Models for War: Inside Its Bootcamp

Overview

Scout AI, a defense-focused startup founded in 2024 by Colby Adcock and Collin Otis, has raised $100 million in Series A funding led by Align Ventures and Draper Associates. The company is building AI agents that control autonomous military vehicles, starting with logistics and moving toward autonomous weapons systems.

Key Technology: Vision Language Action (VLA) Models

Scout AI is leveraging Vision Language Action models — a newer approach to autonomy that builds on large language models (LLMs) to control robots and autonomous vehicles.

  • VLA Origins: First released by Google DeepMind in 2023, VLAs have been adopted by robotics startups like Physical Intelligence and Figure AI
  • Military Application: Unlike traditional autonomous systems, VLAs can operate in unpredictable, off-road environments without structured rules
  • Training Approach: Scout trains its "Fury" model by having human drivers operate ATVs on challenging terrain, then uses reinforcement learning to improve the model based on human takeovers

The "Fury" Model and Training Process

Real-World Training Ground

  • Scout operates at an undisclosed U.S. military base in central California
  • Uses four-seater all-terrain vehicles on unmarked, challenging hillside trails
  • Operations team (former soldiers) conducts 8-hour training shifts
  • Only 6 weeks of training so far on military ATVs

Observable Capabilities

  • Vehicles adapt lane positioning based on trail width (hugging right on wide trails, centering on narrow ones)
  • Faster acceleration than human drivers
  • Sudden slowdowns when the model needs to "think" through complex decisions
  • Successfully completed 6.5 km autonomous loops during demonstrations

Product Stack

1. "Ox" Command & Control Software

  • Bundled with hardened hardware (GPUs, communications, cameras)
  • Allows individual soldiers to orchestrate multiple drones and autonomous ground vehicles
  • Uses natural language prompts like "Go to this waypoint and watch for enemy forces"
  • Expected to be the first widely adopted product

2. Autonomous Weapons Integration

  • Testing drone swarms with a "quarterback" platform providing compute resources
  • Munition drones can search geographic areas and attack targets
  • Programmable constraints: geographic boundaries, human confirmation requirements
  • Vision language models enable better threat identification and targeting

Military Contracts and Adoption

  • $11 million in military technology development contracts from DARPA, Army Applications Laboratory, and other DoD customers
  • One of 20 autonomy companies being tested by U.S. Army's 1st Cavalry Division at Fort Hood, Texas
  • Expected deployment with the unit in 2027
  • Participating in autonomous resupply convoys and security patrols

Technical Architecture

Foundation Models

  • Built on top of existing LLMs from "very well-known hyperscalers" (specific providers not disclosed)
  • Uses pretrained models specialized on military data
  • Full stack includes VLAs, deterministic systems, and other AI flavors
  • Plans to build proprietary model from scratch using new capital

Hypothesis: Path to AGI

  • CTO Collin Otis believes constant real-world interaction could enable Scout to reach AGI faster than internet-trained models
  • "Most intelligence comes with interacting in the world" — not just reading data

First Use Cases: Logistics Over Combat

Automated Resupply:

  • Carrying water/ammunition to distant observation posts
  • Convoy operations: one crewed truck followed by 6-10 autonomous vehicles
  • Saves human labor for higher-value tasks

Future Applications:

  • Reconnaissance and defense
  • Autonomous weapons targeting (with human oversight options)
  • Drone swarm coordination

Strategic Positioning

Mission-Driven Defense Tech

  • Openly critical of tech companies reluctant to work with military (e.g., Google pulling out of Pentagon drone contest, Anthropic's DoD tensions)
  • Positions itself as software company, not vehicle manufacturer
  • Building intelligence layer for existing military assets

Capital Deployment

  • Majority of $100M will fund model training and compute costs
  • Expects to reach scale faster than commercial robotics companies because "customer has assets" (Pentagon's existing vehicle fleet)

Key Differentiators

  • Real-world training emphasis: Unlike simulation-heavy approaches, Scout prioritizes actual terrain and conditions
  • VLA-first architecture: Betting on newer AI approach vs. traditional autonomy systems
  • Military-specific pretraining: Models trained on defense-relevant scenarios from the start
  • Rapid deployment timeline: Only 6 weeks of ATV training already showing promising results