// AUTONOMY WITHOUT GPUS // GPS-DENIED // COMMS-DENIED // 1 WATT // 100 HZ // $10–$50 PER BRAIN // SWARM-NATIVE // ATTRITABLE BY DESIGN // ZERO-SHOT TRANSFER // NO PILOT // NO LINK // NO PROBLEM // AUTONOMY WITHOUT GPUS // GPS-DENIED // 1 WATT // 100 HZ //
STATUS ● ARMED-READY GPS NOT REQUIRED UPLINK NOT REQUIRED
AntBrain // Autonomy for Missiles & Drones
The $10 brainthat fliesitself.
Everyone builds cheap drones. Nobody has made cheap autonomy work. AntBrain runs full flight control, target tracking and swarm behavior on a Raspberry Pi at 1 watt. No GPU, no GPS, no pilot and no link.
The modern battlefield wants thousands of attritable drones. The industry answer is a GPU in every airframe, a pilot on every stick and a datalink that the enemy jams on day one. That doesn't scale. You can't mass-produce it, and it falls apart once the link is cut.
AntBrain makes the brain the cheapest part on the airframe.
Indicative small-UAS pricing. The GPU stack costs more than the airframe it rides on. AntBrain costs less than the props.
The Brain
A predator's nervous system. Not a chatbot.
An ant hunts with about 250,000 neurons. A dragonfly catches prey in midair 95% of the time. Neither one needed a data center. AntBrain uses 4D temporal-spatial recurrent networks that learn time and physics from direct experience. It predicts where the target will be and keeps hunting after line of sight is lost.
~1MParameters total
52×52Pixels is enough to fly
10,000+Step memory horizon
9Discrete flight actions
HoursTo retrain for a new mission
Zero-shotTransfer to unseen terrain
RNN CORE // 7 MODULES NEURONS 112 · RECURRENT
GPU NONE DRAW 1.0 W · LOOP 100 HZ
Nobody else has cracked this — nobody else has cracked this — nobody else has cracked this — nobody else has cracked this —
KillerAnt // Guidance Autonomy
It doesn't follow a waypoint. It hunts.
KillerAnt puts AntBrain in the guidance seat of a missile, interceptor or one-way attack drone. It handles elevator, aileron, rudder and throttle in closed loop, directly from pixels.
01
PURSUIT
Moving-target strike
Coordinates all flight surfaces in real time to run down and hit maneuvering targets.
02
PREDICTION
Hunts through occlusion
Predicts the target's path after it disappears and searches where it will reappear, not where it was last seen.
03
TERRAIN
Clutter at speed
Avoids buildings, terrain and obstacles in real time inside environments it has never seen.
04
DENIED
No GPS. No link.
Uses only a camera and a 6-axis IMU. There's no satellite to spoof and no uplink to jam.
05
TASKING
Point. Click. Strike.
Retask in the field from an image and a text instruction, changeable in real time.
06
ADAPT
Outlearns countermeasures
Persistent belief state with no episodic resets. It keeps adapting for the whole sortie.
Drone Hunting // Urban Intercept
A $50 brain that hunts a Shahed through a city.
Cheap one-way attack drones fly low between buildings. Line of sight breaks every few seconds, and GPS and the radio link are jammed. Here an AntBrain interceptor chases one through a dense urban grid using only its own camera. When the target disappears behind a building, it predicts where it went and keeps hunting.
INTERCEPTOR // AIRFRAME MESH 30K TRIS
BRAIN ANTBRAIN COMPUTE PI-CLASS · 1 W
TOP-DOWN // URBAN GRID MODE SEARCH
RANGE ---- M GPS NONE · LINK NONE
KILLS 0 SENSOR 52×52 CAM + IMU
01
Search
Sweeps the streets with its onboard camera. No external cueing.
02
Track
Locks on, leads the target and closes in.
03
Occluded
The target breaks line of sight. AntBrain projects its path (cyan).
04
Reacquire
Goes to where the target will be, not where it was last seen.
05
Kill
Intercept. Then on to the next target.
AntBrain interceptorShahed-type one-way attack droneTrue position, unseen by interceptorPredicted pathIllustrative simulation of behaviors reported in AntBrain's FastRocketEnv tests. Not flight footage.
Swarm // Colony Intelligence
Kill one. Kill a hundred. The colony still wins.
There's no central controller to decapitate and no master node to jam. Each drone makes its own local decisions, and colony-level tactics emerge from the group. At $10–$50 per brain, you can afford to lose most of the swarm.
AGENTS 0 · LOST 0 · MISSION ON TARGET
The Numbers
Head-to-head. Not close.
Metric
AntBrain
GPT-4 vision
Margin
Mission success
92.3%
43.3%
2.1×
Decision speed
157.8 FPS
0.2 FPS
789×
Compute cost
$200
$181,610
908×
100k-episode eval
4.06 hrs
~188 days
1,113×
Multi-step planning (5-block)
99.72%
—
TRL 6
Partially observable 8×8 gridworld, 50-step cap, 3×3 vision window, run on a $200 Apple M1 CPU. Results come from internal evaluations across six simulated environments, including FastRocketEnv missile guidance and NVIDIA Isaac Lab 3D physics. Self-assessed at TRL 6. Hardware flight validation is the next milestone.
✕
GPS DENIED
It navigates by vision and inertia. There's no satellite to spoof.
✕
COMMS DENIED
Fully onboard decisions. A jammed link changes nothing.
✕
NO PILOT
No operator per airframe. One commander can run a whole colony.
✕
NO DATA CENTER
CPU-only, trained in hours, and fits in a $35 board.
SOCOM // MOD // Primes // Integrators
Mass. Speed. Autonomy.
Swarms need cheap autonomy, and AntBrain delivers it on a $35 board. We're partnering with integrators and defense programs now to take AntBrain from simulation to the airframe.