Case Study: Improving Army Explosive Ordnance Disposal Training with Katana (XR + Offline Computer Vision)
Dauntless XR implemented computer vision-enabled mixed reality training, called the Katana app, for the U.S. Army to support Joint Service Explosive Ordnance Disposal (EOD) training on large-scale, scarce assets, combining rapid 3D digitization with offline computer vision on Meta Quest.
Key outcomes:
Implemented a rapid digitization pipeline for scanning large-scale equipment assets under limited time access windows.
Delivered hands-free guidance and training on the scanned asset Meta Quest using Katana XR
Enabled subject matter expert-authored workflows and analytics review through Katana Pro (web)
Deployed edge-first computer vision for offline object detection on standalone headsets
Supported an asset-agnostic approach by integrating both 3D models from scanned field assets and commercially purchased 3D models
Object Detection & Mixed Reality Project Overview
This implementation applied the Katana platform to Joint Service EOD training and operational field assistance. The goal was to deliver a scalable, hardware-agnostic workflow that supports:
Rapid digitization of large-scale weapons systems and machinery
Edge-based computer vision (CV) on standalone XR headsets
Repeatable training that does not depend on physical asset availability
The work focused on practical constraints found in real military operations: limited access to equipment, the need for offline operation, and performance limits on untethered headsets.
The Challenge: Soldier Training on Scarce and Large-Scale Assets
EOD training requires high-fidelity interaction with specific foreign and domestic systems. In practice, access to the right EOD assets is often limited, and hands-on repetition can be impractical.
Key constraints included:
Restrictive access windows: Critical systems, such as tanks, may be available only briefly. This requires a rapid scanning workflow that prioritizes speed of capture over traditional, labor-intensive 3D modeling.
Asset scarcity and logistics: It is difficult to provide global units with sufficient practice when only 1 to 2 physical items exist for an entire service branch.
Equipment risk: Trainee practice can damage expensive, sensitive, or high-consequence equipment.
Program rigidity: When physical asset availability is delayed, teams still need a way to keep training development moving by using COTS 3D assets alongside proprietary scans.
These constraints are not unique to defense. Similar patterns show up in aviation maintenance, heavy equipment service, energy, manufacturing, and any training program where assets are expensive, distributed, or unsafe for repeated hands-on reps.
The Solution for Computer Vision-Enable Mixed Reality: Katana Pro (Web) and Katana XR (Meta Quest)
The Dauntless XR solution centered on a hardware-agnostic, no-code ecosystem designed to empower Subject Matter Experts (SMEs).
Katana Pro (web authoring and analytics): A browser-based tool where SMEs author and edit workflows, attach 3D assets, and review analytics to improve training content over time.
Katana XR (Meta Quest application): The end-user application that delivers hands-free, in-context guidance on standalone headsets such as Meta Quest 3. It uses spatial anchoring so digital instructions stay aligned to the physical asset or its virtual twin.
Operational mobility: The platform supports untethered use for field conditions and is designed for interoperability across head-mounted displays (HMDs) using open standards.
Mixed Reality: To reduce the chances of “sim sickness” we deployed the solution in mixed reality. Using this medium allows users to see their real world surroundings and immersive, 3D assets without compromising situational awareness.
Rapid Digitization: 3D Scanning Methodology
To support limited access windows and large-scale assets, the Dauntless XR implemented pipelines that capture quickly, then optimize for deployment on standalone headsets.
3D Mesh vs Gaussian Splats
Dauntless XR compared 3D mesh and Gaussian splat technologies to determine the best fit for EOD requirements.
Criteria
3D Mesh
Gaussian Splat
Data type
Geometric (vertices, faces)
Point-based (color, opacity)
Visual quality
Good; depends on poly count
Very high; photorealistic
File size and performance
Can be large; requires decimation
Generally lighter; efficient in XR
Interaction level
Full (measurement, physics, collision)
Limited (view only; no collision)
Best use case
Field mode; precise spatial tasks
VR training; immersive walkthroughs
In practice, both representations have a role. Meshes support interaction and spatial tasks, while splats can provide high visual fidelity for walkthroughs and recognition training.
Why Optimization Is Mandatory on Standalone Headsets
Internal testing on Meta Quest 3 and HoloLens 2 showed that high-fidelity models, such as a 4K aircraft cockpit, can cause frame lag and visible tearing. That degrades user experience and situational awareness.
As a result, the Katana pipeline includes a mandatory optimization step to process raw scans into XR-ready models that run smoothly on untethered mobile chipsets.
Edge-First Computer Vision and Synthetic Data
To provide real-time object detection without relying on persistent connectivity, the implementation used an edge-first computer vision approach.
Model Selection
Ultralytics YOLO v8 Nano was selected for its balance of inference speed and accuracy on edge devices.
Synthetic Data Generation
Real-world EOD imagery is scarce, and manual labeling is time-intensive. Dauntless XR developed an in-house Unity editor extension to generate synthetic training data.
Generated 800,000 labeled images at 640 x 640 resolution to match YOLO input requirements
Used controlled variation to support generalization, including lighting, angles, and occlusion
Data Control and Security
The in-house pipeline supports full control over training data and avoids dependencies on third-party AI tooling.
Compute Resources
Training was localized on a Dell Precision 3680 workstation to maintain data security.
Technical Performance Metrics
Inference speed: 12 to 14 ms per image on-device
Training intensity: 500 epochs for the generalized beta model
Confidence threshold: 70% to balance detection sensitivity and false positives
Outcomes and Success Metrics
The implementation delivered validated performance across scanning, deployment, and offline CV.
Key results:
AI detection accuracy: Achieved a detection rate exceeding 90% on Meta Quest 3 using the synthetic-trained beta model, proving the viability of ONNX model format integration.
Asset flexibility under schedule pressure: When physical access to the T-72 was delayed, the team pivoted to procuring and integrating COTS 3D models, validating the platforms asset-agnostic design.
Workflow delivery: Integrated T-72 startup procedures into Katana XR and validated spatial anchoring for EOD tasks.
Whats Next
The next phase builds on this implementation to support broader deployment and operational hardening across:
Secure 3D asset ingestion and governance
Direct model uploads in Katana Pro
Execution of custom CV models within Katana XR on target headsets
Wider user testing and UX iteration
SOPs for asset management and training integration packages
Katana provides a secure, scalable approach to building a digital library of assets for hands free training and operational support that is not limited by physical availability. By combining rapid digitization with edge-based computer vision, teams can support repeatable training and hands-free guidance in environments where connectivity and access are constrained.
Try Katana XR
Private beta access is currently available for Meta Quest 3 and 3S users in the US Army. To try Katana XR, contact lorilee@dauntlessxr.com with the email address associated with your Meta account to request access.
If you are evaluating Katana for an enterprise XR program, including workflow authoring, offline CV, or asset digitization pipelines, contact us for a technical walkthrough and integration discussion.
