In contested electromagnetic environments, the ability to detect, classify, and locate RF emitters is a core signals intelligence (SIGINT) capability. Operators need to rapidly answer: What's transmitting? Where is it? Is it friendly?

Your task: build Find My Force — a system that takes RF signal data, classifies emitters by type, estimates their geographic location from multi-receiver observations, and presents the tactical picture on a Common Operating Picture (COP). You will work with real IQ waveform data representing common radar and communications modulation techniques, a live simulation feed from a network of simulated receivers, and the challenge of distinguishing friendly signals from hostile and civilian emitters.

Event Date: March 7th, 2026
Event Time: 8:30 AM – 7:30 PM
Build Time: 9:55 AM – 4:00 PM
Format: Software-only; all work is done on your team's laptops using provided datasets and a live simulation feed
Team Size: Up to 5

You are a signals intelligence operator supporting a joint exercise. Multiple friendly platforms — UAVs, ground vehicles, communications relays — are operating in an area alongside civilian RF activity and potential adversary emitters. Your sensor network consists of several receivers at known positions, each reporting RF detections with signal characteristics and received signal strength.

You have been briefed on what your own forces' signals look like — their modulation types, waveform characteristics, and expected signal profiles. You have labeled training data for these friendly emissions. But you have no labeled examples of hostile emitters. Intelligence suggests adversary forces are operating radar and communications systems in the area, but you don't know exactly what they look like. Your system must figure that out.

Your job: build a pipeline that classifies detected signals, locates them geographically, determines whether they're friendly, civilian, or potentially hostile, and displays the full picture so an operator can make decisions under pressure.

This challenge has two core technical problems that feed into each other:

Given a snapshot of raw IQ (in-phase/quadrature) data from an RF emission, classify the signal by modulation type and/or signal type. This is the machine learning core of the challenge.

You will receive a labeled training dataset of real radar and communications waveforms containing several signal types that friendly forces operate. Each sample is a 256-element vector: 128 in-phase components followed by 128 quadrature components, captured at 10 MS/s across a range of signal-to-noise ratios (SNR).

The training data covers friendly signal types only. The live simulation feed will contain all signal types — including hostile emitters and civilian devices that you have never seen labeled examples of. Your classifier must:

• Correctly identify known friendly signal types
• Detect signals that don't match any known friendly pattern and flag them as unknown/hostile
• Handle low-SNR conditions where signals are buried in noise
• Produce a confidence score for each classification

This is a semi-supervised learning problem. You know what friendly looks like; you must discover what hostile looks like. Teams that treat this as a pure supervised classification task will miss the hostile emitters entirely.

When an emitter transmits, multiple receivers in the network may detect it near-simultaneously. But the feed gives you a flat stream of independent receiver observations — it doesn't tell you which observations came from the same emission. Before you can geolocate anything, you need to associate observations into groups that likely correspond to a single emitter at a single moment.

This is an association problem. Consider: if receiver A and receiver B both report a detection within 100ms of each other, with similar IQ characteristics, they probably detected the same emitter. But if three emitters are active at once, you might have 15 observations to sort into 3 groups — and some receivers may not have detected all emitters.

Once you've grouped observations, estimate the emitter's geographic location. Each receiver reports RSSI (which attenuates with distance) and potentially time-of-arrival (which increases with distance). With distance estimates from 3+ receivers at known positions, use trilateration, multilateration, or optimization-based approaches to compute a position fix.

This is a well-defined estimation problem, but it gets interesting when:

• Your association logic is imperfect (wrong groupings lead to wrong positions)
• Some receivers miss the detection (fewer than 3 reports for a given emitter)
• RSSI measurements are noisy or affected by terrain/obstructions
• Multiple emitters are active simultaneously and must be distinguished
• You need to maintain a position track over time as the emitter moves

• Entrepreneurial Drive – 10 Points
  Assessment of the team’s initiative, problem understanding, and overall execution mindset. Judges will evaluate leadership, proactiveness, and the ability to translate ideas into actionable outcomes.
• Algorithm Approach – 5 Points
  Evaluation of the logic, methodology, and structure of the algorithmic solution. Judges will consider originality, efficiency, clarity of design, and appropriateness of the selected approach.
• Technical Implementation – 10 Points
  This is the score from the website. Assessment of the technical quality of the solution, including functionality, robustness, integration, performance, and overall execution. Working prototypes and demonstrated technical competence will score higher.
• Feasibility & Scalability – 5 Points
  Evaluation of the practicality of implementation in real-world scenarios, including operational viability, scalability potential, and resource considerations.
• Storytelling & Vision – 3 Points
  Assessment of clarity, persuasiveness, and strategic vision during the presentation. Judges will evaluate how effectively the team communicates the problem, solution, impact, and long-term vision.