DARPA Wants Drone Video Cut by 99 Percent — Running on Two Watts. DV019's $2M Semantic ISR Topic Closes September 23.

August 24, 2026 · 6 min read

Granted Research Team · Editorial policy

Most compression problems are solved by throwing away detail. DARPA's Semantically-Aware ISR topic asks for something harder: throw away the wrong 99 percent and keep the right one percent, where "right" is defined by a mission intent the operator typed in natural language ten minutes ago.

DPA26BZ05-DV019 opened August 26, 2026 in DARPA's FY26 SBIR Release 5 and closes September 23, 2026 at 12:00 PM ET. It is Direct-to-Phase-II only.

ComponentAmountDuration
Base periodUp to $1,500,00018 months
OptionUp to $500,0006 months
TABA (commercialization assistance)Up to $25,000At award
TotalUp to $2,000,000Up to 24 months

The problem DARPA is actually solving

Small drones generate far more video than tactical networks can carry. The standard mitigations are all unsatisfying in a contested environment: lower the bitrate and you lose the detail that made the sensor worth flying; buffer and forward later and you lose the timeliness; relay to a cloud processor and you have created a high-bandwidth reachback link that is the first thing an adversary jams.

DV019 forbids the third option outright. No cloud processing. No continuous high-bandwidth reachback. Whatever decides what to transmit has to run on the aircraft.

This is the "semantic communications" thesis moving from academic literature into a procurement line item. The academic version has been building for several years — the argument that transmitting mission-relevant abstractions rather than raw multimodal streams yields order-of-magnitude bandwidth reductions, with published results in the high-80s percent range for autonomous vehicle networks. DARPA is asking for 90 to 99 percent, in ISR video, on hardware you can hang off a Group 1 or Group 2 airframe.

Four capabilities, and one of them is the hard one

The topic names four required capabilities:

Mission-intent grounding. An operator states, in natural language, what matters on this sortie. The system converts that into executable logic — without mid-flight retraining. This is the discriminator. A model fine-tuned overnight on today's target set is not a solution; the whole point is that intent changes between sorties and sometimes during them, and the aircraft cannot phone home for new weights.

Spatiotemporal reasoning. The system must recognize mission-relevant behavior, including from objects that are unknown or camouflaged. Note the framing: behavior, not classification. A classifier that has never seen the object cannot label it, but a system reasoning over motion and context can still flag "that thing is doing something that matters."

Multimodal fusion. Electro-optical video plus additional sensor streams.

Auditability. Full traceability for every transmission and suppression decision. If the system chose not to send something, there must be a record of why.

That last requirement is not bureaucratic garnish. In a system whose entire function is deciding what a human never sees, suppression is the failure mode with consequences. DARPA is requiring the audit trail up front rather than discovering later that no one can reconstruct why a frame was dropped.

The power budget is the real gate

DARPA names the hardware: NVIDIA Jetson Orin Nano, Hailo-8L, Coral Edge TPU, AMD Kintex FPGAs, or ARM Cortex-M. And it names the envelope: 5 watts interim, 2 watts final.

Two watts is severe. It is well below what a Jetson Orin Nano draws under sustained inference load in its ordinary configurations, which means a compliant solution is not "run a vision-language model on the edge." It is aggressive quantization, sparse or event-driven inference, hard duty-cycling, or a hybrid where a very cheap always-on stage gates an expensive stage that runs rarely. The Cortex-M appearing on that list alongside a Jetson is DARPA signaling that it expects a tiered architecture, not one model.

If your proposal's power section is a paragraph, you will lose to someone whose power section is a table with measured numbers. This topic will be won and lost on watts, not on accuracy.

Read the exclusion list. It is the strategy.

Buried in the topic is a sentence that tells you more about DARPA's institutional posture than the technical requirements do:

Weapon-release, lethal engagement, and named-person identification are explicitly excluded.

Three separate prohibitions, and they are doing distinct work.

Weapon release and lethal engagement are excluded because DARPA is scoping this as a communications and sensor-management problem, not an autonomy-in-targeting problem. A system that decides what video to send is governed by one review pathway; a system that participates in an engagement decision is governed by a considerably heavier one. Keeping the scope clean keeps the program moving.

Named-person identification is the more interesting exclusion. Facial recognition would be an obvious way to satisfy "mission-relevant" — and DARPA has ruled it out. That constraint pushes the entire solution space toward behavior and pattern reasoning rather than identity matching, which is consistent with the spatiotemporal-reasoning requirement and with the emphasis on unknown and camouflaged objects.

For proposers, this is directly actionable: do not propose an identity-based approach and expect to negotiate scope later. If your existing product line is built on face or person re-identification, the reusable portion is the pipeline and the edge-optimization work, not the model. Say so explicitly in the proposal rather than letting an evaluator wonder.

Who can actually bid

Direct-to-Phase-II requires documented prior feasibility. DV019 specifies the bar as TRL 4 or higher, achieved outside SBIR funding.

That is a real constraint and a slightly unusual one. It is not enough to have done the work — the work must have been funded by something other than a prior SBIR award. Internal R&D, other-transaction agreements, commercial revenue, and non-SBIR government contracts all qualify. A prior Phase II on adjacent technology does not.

The bidder profile this produces is narrow: companies that have already fielded edge-inference video systems commercially, or that have non-SBIR defense work in tactical ISR, or that spun out of a research group with published semantic-communications results and can document lab demonstrations. If you are in that set, four weeks is enough time to write. If you are not, DV019 is a topic to study for Release 6 rather than to bid.

Practical mechanics

Why this topic matters beyond the $2 million

DV019 is a small award attached to a large architectural bet. The Department has been pushing sensor-to-shooter timelines that ground-relay architectures structurally cannot meet in contested environments, and every proposed fix — on-orbit processing, mesh autonomy, edge inference — converges on the same principle: move the decision to where the data is, because the link is the vulnerability.

Semantic ISR is that principle applied to the cheapest, most numerous sensor platform in the inventory. If it works at two watts on a Group 1 drone, the same approach scales to every constrained-link sensor the Department fields — and the commercial parallels in industrial inspection, wildfire monitoring, agricultural survey, and maritime domain awareness are immediate, since all of them involve expensive video pipes carrying mostly nothing.

The winner of DV019 is not selling a compression product. They are qualifying as the supplier of a component that a lot of future programs will need. Price the proposal accordingly, and make the Phase III commercialization narrative carry real weight — DARPA is buying an architecture, and it wants to know the architecture will survive the end of the SBIR.

Get AI Grants Delivered Weekly

New funding opportunities, deadline alerts, and grant writing tips every Tuesday.

Browse all DARPA grants

More DARPA Articles

DARPA Wants AI That Invents Algorithms to Become Ordinary Software. SPEED DIAL Is a $2M Direct-to-Phase-II STTR Closing September 23.

DARPA's DIAL program already had AI rediscover the Kalman filter and wavelets from scratch. SPEED DIAL (DPA26TZ05-DV003) funds the unglamorous next step: turning algorithm discovery into a tool that plugs into MATLAB, Simulink, and COMSOL. Up to $750,000 base plus a $1.25 million option, Direct to Phase II only, with a three-accomplishment feasibility gate and a mandatory university partner. Here is the milestone schedule, the interpretability requirement, and who is actually eligible.

Read article

DARPA's DICE Program Bets $Millions That the Future of Military AI Is a Swarm That Governs Itself — Proposals Close August 25

DARPA's DICE program (HR001126S0010) wants the theory and algorithms for a self-organizing collective of AI agents that stays on mission in contested environments with no central controller. It is a 36-month, simulation-only, three-technical-area BAA due August 25, 2026 — and it rewards a rare blend of distributed-systems theory and frontier-model inference control that almost no single lab has under one roof. Here is what DICE is really asking for, who is positioned to win, and how to build a team that can.

Read article

DARPA's Release 5 Opens August 26 and Closes September 23 — but the Real Advantage Starts August 5, and Almost Nobody Uses It

DARPA's FY26 SBIR Release 5 pre-releases August 5, 2026, opens August 26, and closes September 23 — a four-week window. The Department of War now pre-releases SBIR/STTR topics the first Wednesday of every month, turning DARPA solicitations into a predictable annual cadence you can plan around. The most valuable and least-used part is the three-week pre-release period, when you can talk directly to the topic authors before proposals open. Here is how to turn the monthly rhythm and the pre-release window into a real competitive edge.

Read article

Not sure which grants to apply for?

Use our free grant finder to search active federal funding opportunities by agency, eligibility, and deadline.

Find Grants

Ready to write your next grant?

Draft your proposal with Granted AI. Professional members win a grant in 12 months or get a full refund.

Backed by the Granted Guarantee