Anduril Machine Learning: How Lattice Finds the Drone

"It's a question between smart weapons and dumb weapons." - Palmer Luckey (Founder, Anduril)

A radar sees a dot. A seagull, a hobby quadcopter and a one-way attack drone can all be the same dot. Whoever tells them apart fastest wins the contract.

This week we're talking about Anduril and the machine learning behind modern counter drone technology, and yes, this can get you a very comfortable job.

Here's the idea: Anduril is a defense technology company that builds autonomous systems, including sensor towers, interceptor drones and the Fury combat aircraft, on top of one software platform called Lattice. Lattice fuses radar, camera and radio sensors into a single picture and uses machine learning to detect, track and classify what it sees, so one operator can supervise many machines.

The money says this is where applied ML hiring is heading. Anduril doubled revenue to $2.2 billion in 2025, raised $5 billion at a $61 billion valuation in May 2026, and was reported in July to be in talks at around $100 billion. In March the U.S. Army signed a 10-year enterprise agreement worth up to $20 billion, and the first $87 million task order made Lattice the Army's tactical command-and-control system for counter-drone operations. In July the first production Fury rolled off the line at Arsenal-1, the Ohio factory built to make up to 150 a year.

Strip away the hardware and the core loop is textbook ML engineering under brutal constraints. Birds outnumber drones by orders of magnitude, so accuracy is meaningless. The camera is not pointed at every track, so your best feature is missing exactly when the target is far away. The answer has to come back in well under a second on a computer bolted to a tower. And a probability of 0.9 has to mean 0.9, because a human is about to commit an interceptor on it.

What does Anduril do?

Anduril builds autonomous defense hardware and the software that runs it. Its products include Sentry surveillance towers, Roadrunner and Anvil interceptors, Ghost and Altius drones, and the Fury collaborative combat aircraft. Every product plugs into Lattice, the AI platform that fuses sensor data, classifies targets and lets one operator supervise many autonomous systems. Anduril is privately held.

How does counter drone technology work?

Counter drone systems run four stages. Detect: radar, EO/IR cameras, radio-frequency sensors and microphones flag candidate objects. Track: a filter links detections over time and estimates each object's position and velocity. Classify: a model decides whether a track is a drone, a bird, an aircraft or clutter. Defeat: jam, spoof or intercept, with a human approving the engagement. Machine learning lives in the middle two stages.

The workhorse tracker is still the Kalman filter. It predicts where the object should be from its last velocity, then blends in each new measurement using a gain that depends on how much it trusts the prediction versus the sensor: $$K=P{H}^{\top }{\left(HP{H}^{\top }+R\right)}^{-1}$$ Here P is the uncertainty of the prediction, R is the sensor noise and H maps the state to what the sensor measures. A noisy sensor gives a small gain, so the track follows the prediction. A precise sensor gives a large gain, so the track snaps to the measurement. Every radar tracker in production is some elaboration of that one line.

Classification is where the interesting features live. Spinning rotor blades modulate the radar return with a micro-Doppler signature that birds do not have. Drones hover and turn in ways birds and aircraft do not. The EO/IR camera adds a vision classifier score when it is pointed at the track, and nothing when it is not. Birds are the hard negative: they are the same size and speed as small drones, and there are far more of them.

Anduril vs Palantir: what is different about the machine learning?

Now I'm going to explain what this means for you:

The Skill Employers EXPECT

Learn tracking and classification under heavy imbalance. Whether you are aiming at Anduril, Shield AI or a counter drone startup, the baseline is the same. You should be able to:

• Implement a Kalman filter predict and update step and explain what the process noise and sensor noise control
• Engineer speed, acceleration and turn rate from raw track positions with finite differences
• Train a classifier where positives are a few percent of tracks, and report PR-AUC instead of accuracy
• Handle a sensor that is missing for a reason (the camera was not tasked) without imputing the signal away
• Write fast, clean Python and know which parts would move to C++ on an edge computer

If your drone detector's headline number is accuracy, you have built a bird detector.

The Skill That Separates You

Learn to design for the operator, not the leaderboard.

One operator watching many sensors can only absorb so many alerts an hour. Strong engineers build the model around that person. You should know how to:

• Set an alert budget and pick the threshold from it, not from 0.5
• Price the threshold: a missed drone against a wasted operator minute
• Calibrate probabilities so a 0.9 is right about 90% of the time, and check it with a reliability diagram
• Score a whole sweep of tracks inside a latency budget on modest hardware
• Test against distribution shift: new drone types, bad weather, a degraded sensor

The Project To Learn These Skills This Week

Build the drone classifier an operator would actually trust. This is a graded Pro project on GRADuateML: Counter-UAS Track Triage: Drone or Bird?. You get 48,000 synthetic radar and camera tracks where only 7.9% are drones, birds with drone-sized radar returns, a camera that looked at only 28% of tracks, a hidden test set, and instant grading on every milestone. The punchline: a model on kinematics alone reaches a PR-AUC of 0.39. Fuse in micro-Doppler and the camera score and it reaches 0.72, and inside a budget of 11.5% of tracks it catches 74% of drones instead of 45%. Not on Pro? The free challenge below is one Kalman filter step, by hand.

Requirements (each one is a graded milestone):

• Read the air picture: the class mix, how often the camera was tasked, and the drone rate with and without a camera look.
• Derive speed and turn rate from raw radar plots by finite differences, wrapping headings correctly (a hovering drone's heading is pure noise).
• Break it on purpose: a system that never alerts scores 92.3% accuracy and catches zero drones.
• Fuse the sensors: PR-AUC of at least 0.62 and ROC-AUC of at least 0.90 on hidden tracks.
• Fit the operator's attention: cue no more than 12% of tracks and still catch at least 65% of drones.
• Price a missed drone at 40 times a wasted operator minute, sweep the threshold on held-out data, and find that the best operating point cues about three times what one operator can absorb.
• Score the whole sweep on the edge node within 3 seconds, training included, then defend your operating point in writing.

This project teaches a practical lesson for production ML:

In the field, the best model is the one that spends a human's attention wisely.

If you want to sharpen your machine learning skills even more, I also selected a free challenge problem for you this week. It is one step of the Kalman filter from above, by hand:

If you learned something from this newsletter, make sure to forward it to a friend.

GRADuateML

The full counter-drone track classification build, graded against a hidden test set, is part of GRADuateML Pro.