The AI Military Academy: Combat-Ready AI Operators
An integral part of 2DefAi.tech's defense technology solutions, focusing on developing and training AI agents to become combat-ready and highly qualified military operators. Creating a new generation of autonomous systems capable of performing specific missions, piloting, and co-operating weapons systems.
AI Military Academy: Training and Development
The academy employs sophisticated methodologies to train its AI agents, ensuring they are prepared for real-world combat scenarios and diverse operational environments.
Hybrid Training Paradigm
Behavioral Cloning (BC)
Provides the initial training, where the AI model learns by imitating "expert demonstrations" from skilled human operators in simulated environments. This establishes a strong, safe baseline of competent behavior, teaching fundamentals like navigation and control.
Reinforcement Learning (RL)
Allows the AI to innovate and improve beyond human performance. After BC, the model interacts autonomously with the simulation, receiving rewards for desirable actions and penalties for undesirable ones. This iterative process refines the AI's policies, enabling it to adapt, recover from errors, and achieve superhuman optimization and adaptability.
Parameter-Efficient Fine-Tuning
PEFT/LoRA techniques are used to make this training computationally feasible, significantly reducing memory and computational requirements for fine-tuning large models like Gemma 3n.
Advanced Training Methodologies
Simulation-Based Learning
AI agents undergo thousands of hours of continuous training and virtualized simulation in environments that replicate real-world combat scenarios. Game developers create highly realistic simulation environments, such as AirSim, customized for specific geographical areas of operations.
Adaptive Learning Protocols
The AI systems exhibit continuous improvement through real-time feedback and mission outcome analysis, enabling dynamic scenario adjustments based on performance metrics.
Expert Knowledge Integration
Veteran special forces consultants play a critical role in designing AI training programs to ensure realism and effectiveness. Human operators also contribute by generating the initial "expert demonstration" datasets.
Specialised Training for Combat Environments
Geo-Specific Training
Training utilises specialised datasets tailored for Ukrainian theatre operations and regional tactical requirements, as well as specific vessel types and mission profiles.
Mission-Specific Training
Customised training regimens prepare AI for particular mission types and operational parameters.
EW-Resilient Operations
AI is prepared to operate effectively even in challenging conditions like heavy electronic warfare (EW), relying solely on visual, maps, and compass data.
The AI's Cognitive Core: Gemma 3n
The project's most significant technological leap is the use of a Commercial Off-The-Shelf (COTS) smartphone (specifically Samsung S24/S25 series) as the complete onboard processing unit, running a sophisticated multimodal AI model.
Gemma 3n as the Unified "Brain"
Google's Gemma 3n (specifically version 3n-E4B) serves as the unified "brain" for all decision-making, navigation, perception, and tactical execution, collapsing the traditional robotics architecture into a single, powerful, on-device model. While LLaMA 2 was initially considered for some aspects, Gemma 3n is specified as the core model for the current system.
Multimodal Input Fusion
Gemma 3n directly processes sequences of video frames from the camera, eliminating the need for separate Convolutional Neural Networks (CNNs) for feature extraction. It uses an optimised MobileNet-V5 encoder. All other telemetry and sensor data (GPS, IMU, speed, battery status, target coordinates) are converted into a structured, human-readable text string, leveraging Gemma's language understanding capabilities and replacing the need for Recurrent Neural Networks (RNNs) for sensor data. This creates a unified prompt that allows Gemma to "see" (video frames) and "feel" (sensor text) the ASV's state.
On-Device Deployment
The fully trained and optimised Gemma 3n model is deployed directly onto the smartphone or specialised edge computing devices (like NVIDIA Jetson) using the Google AI Edge SDK and converted to the TensorFlow Lite (.tflite) format.

This ensures offline operation, independent of external connectivity, which is critical for electronically contested environments.
The phone's native Neural Processing Unit (NPU) and Graphics Processing Unit (GPU) are leveraged for efficient real-time inference.
AI's Operational Roles and Capabilities
The AI agents developed through the academy have a versatile range of roles and capabilities:
1
Autonomous Decision-Making
The AI systems enable independent operational capability, particularly in communications-denied environments, by analysing environmental data, identifying threats or targets, and making tactical decisions based on programmed rules and mission objectives.
Hybrid Mission Execution & Human-AI Cooperation
AI systems assist human pilots through intelligent automation while maintaining human oversight for critical decision-making. This includes two operational modes:
AI as a Co-Pilot
The human is the primary pilot, with AI assisting with navigation, system monitoring, electronic warfare (EW), and target acquisition/tracking.
Human as a Mission Commander
The AI assumes the primary pilot role, executing pre-planned missions autonomously, allowing a single human soldier to command an entire group of autonomous drones simultaneously, crucial for swarm capabilities.
Advanced Operational Capabilities
EW Resilience and Autonomous Navigation
The AI is trained to operate effectively under strong electronic warfare conditions, using only visual, maps, and compass data for navigation. It provides independent pathfinding and obstacle avoidance in complex environments.
Real-Time Decision Support
The AI offers instant tactical recommendations based on battlefield analysis.
Adaptive Mission Execution
AI operators can perform missions based on predefined parameters or improvise as conditions require.
Automated Engagement
The system allows for full automation of target following and engagement under human operator supervision.
Perception and Targeting
A parallel computer vision pipeline runs on the smartphone, using YOLOv8 for real-time target acquisition and efficient tracking algorithms (like CSRT) to identify and track naval targets.
This targeting information is fed into the Gemma 3n decision loop, allowing the AI to execute complex targeting behaviours.

This modularity allows for rapid retraining of target detectors without retraining the core AI model.
Broader Implementations: Swarm Intelligence
A key broader implementation of this AI technology is the development of decentralised swarm intelligence, enabling a fleet of autonomous surface vessels (ASVs) to operate as a cohesive, intelligent, and coordinated unit without relying on a central command-and-control node.
Decentralised Control
This approach ensures robustness, scalability, and adaptability. With no central leader, the system is resilient to the loss of individual units, and its control logic is independent of the number of agents, allowing for effective operation with 10, 50, or even 100 units. Each ASV functions as a self-contained tactical unit.
Communication-Minimised Coordination
In EW-contested environments, coordination is achieved primarily through shared intent via homogeneous training (all ASVs loaded with identical mission plans and trained Gemma 3n models). Opportunistic, short-range, intermittent data exchange can occur when ASVs are in proximity, allowing for local mesh networks to share critical data.
Swarm Algorithms
The project adapts principles from established swarm intelligence algorithms:
  • Ant Colony Optimization (ACO) for coordinated search and area denial (e.g., mine-sweeping).
  • Particle Swarm Optimization (PSO) for coordinated attack formations, allowing the swarm to dynamically converge on optimal, multi-pronged attack vectors.
  • Multi-Agent Reinforcement Learning (MARL): The RL training for Gemma 3n takes place in multi-agent AirSim environments, with reward functions crafted to incentivise cooperative behaviour, leading to emergent team-based tactics.
Safety and Reliability Mechanisms
Given the critical nature of military applications, the AI systems incorporate multi-layered safety mechanisms:
Autonomous Failsafe
Built-in mechanisms activate upon critical system failures or anomalous behaviour, including automatic safe mode, engine shutdown, or return to a last known safe point.
Rollback Mechanisms
If unauthorised or suspicious software changes are detected, the AI system can perform a "rollback" to a previous, verified state.
Emergency Shutdown/Recovery
Protocols allow for remote or conditional activation of emergency shutdown and recovery, preventing the vessel from attacking friendly forces.
Friend/Foe Recognition
Systems integrated with the AI ensure clear differentiation to prevent friendly fire incidents.
Dual AI Supervisor Model
For enhanced reliability, a second, independent AI model acts as a "supervisor," monitoring the operational AI's mission execution, adherence to protocols, and safety limits. If deviations or threats are detected, the supervisor can intervene, initiate safety protocols, or alert human operators.
Broader Implementation Context
This AI development is part of 2DefAi.tech's strategic partnership with the Barracuda, a Ukrainian special forces unit. The project aims to counter numerical superiority by leveraging advanced AI through:
  1. Manufacturing New Units: Developing and mass micro-producing new "Barracuda" autonomous boats, with a goal to expand production from 5 to 50 boats per month using 3D printing technologies.
  1. Robo-Enabling Existing Assets: A novel service model to immediately retrofit existing boats and ground vehicles from other Ukrainian military units with remote control systems and AI capabilities, enabling rapid and large-scale implementation using existing unit budgets for "servicing".
The AI's capabilities are being integrated into the Barracuda fleet, which includes various types of unmanned drone boats and jet skis for diverse critical objectives such as destruction of enemy targets, tactical leadership, strikes on the enemy's rear, mine clearance, reconnaissance, and even delivery of other drones. The project also envisions future commercial potential for private security and other post-war applications.