Resource Library

Drone threats are evolving fast. Drone detection AI needs to evolve faster. This paper presents Duality’s joint work with the U.S. Army Ground Combat Systems Advanced Capabilities Team. It (1) explores how we generate high-fidelity synthetic data for drone detection, (2) details methods for evaluating its quality and realism, and, (3) offers model performance results and insights from real-only, synthetic-only, and mixed-data training. (Awarded Best Paper at GVSETS 2025)

This paper, stemming from Duality's work with NASA-JPL in DARPA’s RACER program, introduces a novel digital twin approach for building off-road autonomy — showing how synthetic data can fine-tune foundation models for semantic segmentation and outperform models trained solely on real-world datasets. By generating synthetic data from photorealistic, physics-accurate 3D environments with digital twins of autonomous vehicles, we demonstrate how domain-specific synthetic data enables more reliable perception training in novel, complex, and variable terrains.

The Task Coding Framework takes advantage of the generative capabilities of LLMs to perform general reasoning to take high level policy goals, framed as prompts, and turn them into well defined coded tasks capable of being directly run on a robotic system. This presents an effective approach to make robotic systems less brittle, with more generalized domains of operations, while keeping their behavior predictable.

Modularity and reusability are key features driving digital twin simulation — so how do we create real interoperability that ensures a meaningful and long lasting digital twin infrastructure? Presented at I/ITSEC 2023, this paper, titled The Digital Twin Encapsulation Standard: An Open Standard Proposal for Simulation-Ready Digital Twins, presents a path forward that ensures full reusability of digital twins in any operational context, while avoiding the possibility of vendor lock-in.

Training a model on poor quality data can cause it to pick up “bad habits” that don’t transfer well to real-world scenarios. So how do we quantify if our synthetic data is relevant and valuable?