DPA27BZ02-DV002: Metacognitive AI for ISR Collection

Part of the DoW SBIR/STTR FY27 Release 2 topic list. Browse all 83 topics by industry or service.

Topic DPA27BZ02-DV002 from the DoW SBIR 2027 BAA, Release 2. Pre-release now; proposals due January 6, 2027, 12:00 PM ET.

Quick answer

DPA27BZ02-DV002 is a DARPA SBIR topic seeking an autonomous, metacognitive AI that decides what intelligence to collect, when, and why across asynchronous multi-domain sensor streams, under hard operational constraints and with auditable decision traces. It is Direct-to-Phase II only, with a 24-month Phase II. Proposals are due January 6, 2027, 12:00 PM ET.

Book a Free Fit Call

Solicitation at a glance

• Topic number: DPA27BZ02-DV002
• Official title: Autonomous Metacognitive Collection Orchestration for Multi-Domain Intelligence
• Agency: DARPA
• Program: SBIR (DoW SBIR 2027 BAA, Release 2)
• Award type: Direct to Phase II (no Phase I awards); up to $1.5M plus a $500K option
• CMMC level: Level 2 (Self)
• ITAR: Yes. Export-controlled (ITAR/EAR); see topic text
• Technology areas: Information Systems
• Pre-release (direct TPOC contact allowed): November 4 to December 2, 2026, 8:00 AM ET
• Opens in DSIP: December 2, 2026, 8:00 AM ET
• Proposals due: January 6, 2027, 12:00 PM ET
• Official source: DSIP Topics app (search DPA27BZ02-DV002)

BW&CO's take

What the DARPA wants

ISR sensors report at tempos that differ by four or more orders of magnitude, and current fusion, reinforcement-learning, and agentic-LLM approaches either assume aligned inputs, trade away constraints, or cannot be audited. DARPA wants a compositional substrate (hyperdimensional, neuro-symbolic, graph-based, or a hybrid) that computes expected information gain for collection actions, treats emissions, bandwidth, power, and time-to-decision as hard constraints, returns Pareto-optimal option-sets with "why / why-not" traces from internal state, and refines itself under deception without offline retraining. Phase II scales from two to four modalities and goes from TRL 4 to TRL 6.

Who is a good fit

• Published or demonstrated HDC/VSA, neuro-symbolic, or graph-based reasoning across at least two heterogeneous data streams with very different update rates
• Value-of-information or sensor-selection planning work (for example, constrained POMDPs or adaptive submodular optimization) that includes power, emissions, bandwidth, or latency costs
• Quantitative side-by-side results against mainstream baselines such as classical fusion, reinforcement or active learning, or agentic-LLM orchestration
• Experience with ISR data types such as EO/IR, SAR, RF, or AIS, and with edge deployment on SWaP-constrained hardware

Watch-outs

• Direct-to-Phase II only: Phase I proposals will not be accepted or reviewed. The feasibility package needs quantitative baseline comparisons, performance under at least 20 percent missing or delayed inputs, video or data artifacts, a risk register, and a transition plan.
• Out of scope: approaches that rely on monolithic end-to-end deep networks without compositional or causal structure, or on narrative-only explanations produced by prompted language models.
• Phase II quantitative goals include at least 35 percent faster time-to-decisive-cue and at least 2x information gain per unit energy (both versus a reactive fusion baseline), constraint violations below 5 percent under adversarial stress, plan refresh under 3 minutes, and at least 85 percent SME coherence rating of traces.
• The topic metadata marks it as ITAR-restricted, so plan for export-control handling of technical data and possible limits on foreign national participation. The metadata lists CMMC Level 2 (Self).

Think your company fits DPA27BZ02-DV002? Talk with BW&CO before the January 6 deadline.

Plan Your DPA27BZ02-DV002 Submission

Official topic text (verbatim from DSIP)

The text in this section is copied word for word from the official DSIP topic as published on November 4, 2026. DSIP is the official source and topics can be amended. Always confirm the current version at dodsbirsttr.mil before you submit. Our commentary above is BW&CO's interpretation, not Government guidance.

Title

Autonomous Metacognitive Collection Orchestration for Multi-Domain Intelligence

Objective

Develop and demonstrate an autonomous, metacognitive artificial-intelligence capability that orchestrates multi-modal intelligence collection across asynchronous sensing streams and produces traceable, causally-informed recommendations for multi-domain operations. The capability shall decide what evidence to collect, when, and why; weigh expected information gain against hard operational constraints including emissions, bandwidth, power, time-to-decision, and access posture; generate Pareto-optimal option-sets accompanied by “why / why-not” traces reproducible from internal system state; and locally refine its internal representation under denial, deception, or distribution shift without offline retraining. Operator interaction is confined to pre-mission intent specification and after-action audit of the decision record, enabling a human-machine formation in which the capability operates at machine speed inside a commander-specified envelope and the human sets intent, constraints, and rules of engagement at mission entry and reviews the auditable decision record at mission exit.

Description

Modern intelligence, surveillance, and reconnaissance (ISR) enterprises ingest multi-source indicator streams whose native tempos span four or more orders of magnitude. Overhead imagery (EO/IR, SAR, MSI, HSI) arrives at revisit cadences of minutes to hours; RF and SIGINT detections resolve in seconds to minutes; maritime AIS and acoustic returns update continuously at sub-second tempos; and open-source, supply-chain, and pattern-of-life indicators mature over hours to weeks. A single decision of operational consequence, such as whether a vessel of interest is a combatant or a civilian fisher, or whether an observed buildup presages a near-term operation or reflects routine activity, typically turns on whether causally related evidence drawn from several of these phenomenologies can be assembled, reasoned about, and surfaced fast enough and cheaply enough to matter.

Classical early- and late-fusion pipelines assume temporally aligned inputs and emit single snapshot outputs, useful when cadences match and brittle when they differ by orders of magnitude or when a depended-on sensor is denied. Reinforcement- and active-learning policies can select next observations but treat operational constraints as reward-shaped scalars the optimizer can trade away, and require retraining cycles measured in weeks when priority intelligence requirements or adversary tactics shift in hours. Neuro-symbolic and causal methods contribute structure and interpretability, yet in practice operate over hand-authored graphs and synchronous data, with post-hoc explanations rather than forward-looking traces. Agentic large-language-model orchestrators produce narrative rationalizations whose internal state cannot be audited, and whose plans cannot be inspected for constraint feasibility before execution. What is missing across all of these approaches is a structured, auditable, constraint-feasible substrate whose rationale for each recommendation is transparent, reproducible, and directly inspectable from internal state.

This topic seeks novel approaches to artificial metacognition: an orchestration substrate that reasons about its own reasoning while reasoning about the world. A metacognitive substrate maintains a structured model of evidence-to-outcome relationships across heterogeneous modalities and tempos; tracks its own uncertainty, prediction error, and representational coherence; computes expected information gain for candidate collection actions under hard operational constraints; returns option-sets rather than single collapsed policies; and, when denial, deception, or distribution shift is detected, locally refines its internal model rather than triggering a full retraining cycle. The resulting capability is the computational backbone of a human-machine formation: the autonomous system acts at machine speed inside a bounded operating envelope, while the human commander sets objectives, constraints, and rules of engagement before the mission and reviews the auditable decision record after.

Candidate substrate families for the metacognitive layer include hyperdimensional computing and vector-symbolic architectures; neuro-symbolic architectures that couple learned encoders with symbolic or probabilistic inference; and graph-based reasoning over typed knowledge graphs, causal graphs, and structured inference engines. Hybrids across these approaches are encouraged. This topic is deliberately substrate-agnostic within that class. Offerors are expected to justify their choice on quantitative grounds. Any viable substrate shall provide:

•  Compositional, inspectable operators (algebraic, logical, graph-based, or probabilistic) that support real-time assembly and decomposition of multi-source evidence structures, rather than end-to-end learned embeddings whose compositional semantics cannot be examined.
•  Bounded degradation under partial, corrupted, delayed, or adversarially manipulated input, quantified against a mainstream deep-learning baseline.
•  Transparent, reproducible reasoning operations that yield auditable explanations directly from internal system state, rather than post-hoc narrative produced by a prompted language model.
•  Efficient operation under sparse and few-shot regimes, with measurable performance at small label budgets on novel scenario classes.
•  A credible path to tactical-edge deployment on Size-Weight-and-Power-constrained platforms.

Recent work across these areas has independently demonstrated subsets of these properties, including graph-structured anomaly reasoning, compositional multi-modal scene understanding, and interpretable edge AI with online adaptation.

Converging them into a single autonomous metacognitive orchestration substrate is the objective of this topic. Approaches that rely on monolithic end-to-end deep networks without compositional or causal structure, or on narrative-only explanations produced by prompted language models, fall outside the scope of this topic.

The program seeks novel approaches to key technical challenges, including:

•  Representing and reasoning over time-lagged causal dependencies across multi-modal sensor streams whose native tempos span four or more orders of magnitude, under missing, late, corrupted, and adversarially manipulated inputs.
•  Computing expected information gain over candidate collection actions while treating operational constraints (emissions, bandwidth, power, time-to-decision, access) as hard-binding semantic variables rather than reward-shaped scalars, and producing Pareto-optimal option-sets rather than single collapsed policies.
•  Self-monitoring the system’s own confidence, prediction error, and representational coherence; detecting adversarial denial, deception, and distribution shift as learning signals; and triggering targeted local refinement of the internal model without offline retraining.
•  Generating forward-looking “why / why-not” traces, reproducible from internal system state, that link each recommendation to its supporting evidence chain, the constraints it satisfies, the alternatives it dominated, and the expected futures it forecasts, supporting downstream review by a subject-matter expert (SME) or within a human-machine formation.
•  Probabilistic root-cause analysis of observed outcomes, tracing back through the internal representation to identify contributing evidence across domains and timescales, with ranked alternative hypotheses and quantified confidence.

Phase I

This SBIR topic is open to Direct-to-Phase-II (DP2) proposals only. Phase I proposals will not be accepted or reviewed. DP2 feasibility shall be substantiated through prior work completed outside of the SBIR program. Acceptable evidence includes peer-reviewed publications, pre-prints, patents, prototype demonstrations with video or data artifacts, technical reports, and test-and-measurement data. Proposers shall demonstrate scientific and technical merit and feasibility across the criteria below, each with quantitative side-by-side comparison against at least one mainstream baseline (classical fusion, reinforcement or active learning, agentic-LLM orchestration, or equivalent) on a shared task.

•  Structured reasoning over asynchronous multi-source evidence. Evidence of a compositional or causally-structured representation, built using hyperdimensional and vector-symbolic architectures (HDC/VSA), neuro-symbolic, graph-based reasoning, or a hybrid, operating across at least two heterogeneous modalities whose native tempos differ by two or more orders of magnitude. The proposer shall characterize the substrate using metrics appropriate to the chosen family (for example, similarity-based retrieval and compositional-capacity metrics for HDC/VSA; inference accuracy and rule-coverage for neuro-symbolic; path-query accuracy and graph-completion metrics for graph-based) and quantify performance under at least 20 percent missing or delayed inputs.

•  Cost-aware, forward-looking action selection. Evidence of an information-gain or value-of-information planning capability, implemented by any appropriate technique including adaptive submodular optimization, constrained partially-observable Markov decision processes, information-theoretic sensor selection, or equivalent, that incorporates at least one operational cost term (power, emissions, bandwidth, or latency) and produces multiple candidate actions or an option-set rather than a single collapsed policy, with quantitative information-per-cost gains over a uniform, random, or reactive baseline.

•  Auditable decision outputs grounded in internal state. Example decision outputs accompanied by traces, reproducible from internal system state, that identify which evidence and which reasoning pathway drove each recommendation and, where applicable, which alternatives were dominated and why.

The DP2 feasibility documentation package shall include a consolidated technical report summarizing the criteria above with quantitative metrics; substantiating publications, reports, or artifacts; video or data artifacts demonstrating operation on at least one multi-domain task with asynchronous inputs; a risk register identifying the highest technical and programmatic risks with mitigations; and a Phase II transition plan.

Phase II

The goal of Phase II is to mature the DP2-validated capability into an operationally representative prototype of an autonomous metacognitive orchestration engine for multi-domain intelligence collection, suitable for integration into a human-machine formation. The prototype shall be demonstrated against at least one mainstream baseline family (classical fusion, reinforcement or active learning, or agentic-LLM orchestration) on shared tasks. The capability shall advance from TRL 4 at award to TRL 6 by the end of Phase II. Phase II is a 24-month period with all required capabilities, targets, and capstone demonstrations completed within this period.

DP2 proposals should:

•  describe a proposal to achieve the aforementioned goals;
•  present a technical plan and approach, with notable risks and mitigations; and
•  detail proposed metrics, baselines, and scenarios for final evaluation, aligned with the quantitative targets.

Staged modality scope. The Phase II effort is staged by modality count. In Months 1 through 12, the prototype shall operate across at least two heterogeneous modalities whose native tempos differ by two or more orders of magnitude, demonstrating the full required capability set on that two-modality configuration. In Months 12 through 24, the prototype shall scale to at least four heterogeneous modalities and shall be evaluated against active adversarial denial and deception, integration with one analytic or data-fabric environment identified during the period, hardening for tactical-edge deployment on Size-Weight-and-Power-constrained platforms in denied, degraded, intermittent, and limited-bandwidth (DDIL) settings, and human-machine-formation integration.

At the end of Phase II the integrated prototype shall demonstrate:

•  Asynchronous multi-source ingest. Streaming ingest from at least four disparate modalities through modular encoder interfaces, with demonstrated tolerance to missing, late, corrupted, and spoofed streams.
•  Compositional reasoning at multiple abstraction levels. A hierarchical structured representation, built in the proposer’s chosen substrate family, that supports primitive (sensor-level), composite (cross-sensor spatio-temporal), and semantic (contextual) abstractions, updated continuously from telemetry.
•  Cost-aware, forward-looking collection planning. Computation of expected information gain over candidate collection actions, discounted by tunable cost terms including at minimum emissions or signature, power, latency, and bandwidth, with generation of Pareto-optimal option-sets rather than single collapsed policies.
•  Auditable decision traces. Forward-looking “why / why-not” explanations for each recommendation, reproducible from internal system state, identifying supporting evidence, constraints satisfied, and dominated alternatives. Narrative-only explanations produced by prompted language models are not acceptable.
•  Real-time re-orchestration without offline retraining. Replanning under sensor loss, intermittent denial, deception injection, or changes in specified objectives and weights, at plan-refresh latency below 3 minutes.
•  Probabilistic root-cause analysis. Tracing observed outcomes back to contributing evidence across modalities and timescales, with ranked alternative hypotheses and quantified confidence.
•  Metacognitive self-assessment and adaptation. Detection of internal reasoning drift, adversarial manipulation, or distribution shift, and targeted local refinement of the internal representation without offline retraining. Preserved or improved performance across at least five repeated exposures to the same class of adversarial perturbation, relative to a baseline that must retrain to recover, is the primary evidence of this capability.
•  Human-machine-formation integration. Modular, open-schema APIs for option-sets, explanations, and constraints, suitable for integration with analytic and data-fabric environments used by a human-machine formation, together with a documented commander-intent interface for pre-mission specification and an after-action audit interface for post-mission review.

Baselines and the capstone scenario set will be finalized with the Program Manager at kickoff. At least one capstone scenario shall involve active adversarial denial or deception. Quantitative goals to be met or exceeded by the end of Phase II include:

•  Time-to-decisive-cue at least 35 percent faster than a reactive fusion baseline at fixed sensing budget.
•  Information-gain per unit energy at least 2 times over a reactive fusion baseline on an energy-instrumented testbed.
•  Operational-constraint violation rate (including emissions-control posture) below 5 percent under adversarial stress.
•  Plan-refresh latency below 3 minutes under induced sensor loss, deception injection, or mid-mission objective change.
•  Subject-matter expert coherence rating of “why / why-not” traces at or above 85 percent against doctrine-aligned rubrics.

The following schedule of milestones and deliverables is provided to establish expectations for the Phase II effort. Proposers shall execute a Research and Development plan as described in their proposal, including the following:

•  Month 1. Kickoff briefing (with annotated slides) to the DARPA Program Manager (PM), including any updates to the proposed plan and technical approach, risks and mitigations, schedule with planned capability milestones and deliverables, proposed metrics and baselines, specification of open-schema interfaces for option-sets, explanations, and constraints, and the planned human-machine-formation integration approach.
•  Month 3. Architecture Design Review. Formal specification of the structured representation, value-of-information planner, and metacognitive self-assessment layer. Baseline benchmarks on one multi-domain scenario involving two modalities at asynchronous tempos and two operational cost terms. Quarterly technical progress report.
•  Month 6. Two-modality prototype demonstration with asynchronous ingest and primitive-level structured reasoning; initial cost-aware planning with two cost terms and option-set generation; quantitative comparison against the reactive fusion baseline on time-to-cue and information-per-cost. Quarterly technical progress report.
•  Month 9. Interim technical progress briefing (with annotated slides) to the DARPA PM. Demonstration of constraint-aware re-orchestration under induced sensor loss and mid-mission objective change; first SME evaluation of “why / why-not” traces against doctrine-aligned rubrics; updated risks, mitigations, and technical plan for the second half of Phase II.
•  Month 12. Two-modality Critical Design Review. Integrated demonstration of the required capability set on the two-modality configuration with asynchronous ingest, meeting or approaching the 35 percent time-to-decisive-cue, 2x information-per-watt, and below-3-minute plan-refresh targets. Documented transition plan from two-modality to four-modality scope. Quarterly technical progress report.
•  Month 15. Four-modality ingest and primitive-through-semantic reasoning demonstration. Quantitative results on representational capacity, noise tolerance, and inference latency on four modalities. First integration with an analytic or data-fabric environment identified during the first half of Phase II.
•  Month 18. Metacognitive adaptation demonstration. Scripted denial-and-deception injection. Quantitative demonstration of performance preservation or improvement across at least five repeated adversarial exposures without external retraining, compared against a baseline that must retrain to recover. Quarterly technical progress report.

Option Milestones

•  Month 21. Human-machine-formation integration and tactical-edge hardening. Full quantitative-target evaluation on at least two capstone scenarios, including at least one with active adversarial denial or deception. Commander-intent and after-action-audit interfaces exercised with SMEs. Quarterly technical progress report.
•  Month 24. Final technical progress briefing (with annotated slides) to the DARPA PM. Capstone demonstration across four modalities on at least three scenarios drawn from the kickoff-agreed set. Final Phase II Report documenting architecture, algorithms, evaluation methodology, quantitative results against all Phase II quantitative goals, API specifications, user and administrator manuals, adversarial-robustness characterization, commercialization plan, and a Phase III transition plan with named non-SBIR funding source(s).

Capstone scenarios. The prototype shall be demonstrated on at least three scenarios drawn from the categories below, agreed with the Program Manager at kickoff. At least one shall involve active adversarial denial or deception.

Scenario 1. Multi-modal collection under emissions control. Orchestrate collection across EO/IR, SAR, RF, and AIS (or equivalent) modalities under a restrictive emissions-control posture with variable spectrum availability. Demonstrate improved information-gain-per-cost against the proposer’s chosen baseline, emissions posture enforced as a hard constraint, bounded degradation under spectrum denial, and recovery when access is restored. Include a scripted injection of a contact-of-interest whose correct classification requires cross-referencing at least three modalities.

Scenario 2. Cross-INT corroboration within a human-machine formation. Support development of a decision product that requires corroboration across at least three independent modalities against a commander-specified hold-fire confidence threshold. Demonstrate that the engine autonomously sequences supporting collection (for example, tasks a revisit to disambiguate an ambiguous cue), returns option-sets at varying confidence, latency, and signature trade-offs, surfaces a reproducible decision trace for downstream review within the human-machine formation, and declines to cross the threshold when corroboration is insufficient, articulating why.

Scenario 3. Denial-and-deception resilience with metacognitive adaptation. Against scripted adversarial injection of coordinated deceptive indicators (spoofed tracks, falsified pattern-of-life, counterfeit logistics signatures, misleading RF emissions), detect semantic incoherence through metacognitive self-assessment and autonomously refine the internal representation to compensate. Demonstrate that performance is preserved or improved across at least five repeated adversarial exposures, measured against a baseline that must retrain to recover.

The awardee shall participate in at least one simulation exercise or demonstration event during the 24-month period of performance, coordinated with the Program Manager.

Phase III Dual Use Applications

Phase III work will be oriented toward transition and commercialization. Funding is obtained from the private sector, a non-SBIR Government source, or both, to develop the prototype into a viable product or non-R&D service for sale in military or private-sector markets. Military applications include autonomous orchestration of persistent multi-domain ISR collection; multi-INT corroboration and decision support for Joint and Combined Joint All-Domain Command and Control (JADC2 / CJADC2) mission threads; cross-domain indications and warning at theater scale; and human-machine-formation decision tooling for multi-domain formations operating under emissions, bandwidth, and access constraints. The capability’s constraint-aware, causally-informed, and auditable reasoning directly supports Department of Defense Responsible AI principles and downstream Operational Test and Evaluation, Mission Assurance, and Authority-to-Operate review processes for AI capabilities that influence collection, tasking, and other consequential decisions. Naval Information Warfare Center Pacific (NIWC Pacific) is a representative candidate Government transition venue. Additional transition opportunities will be identified during execution.

Commercial applications are broad and immediate. The underlying capability (constraint-aware, auditable orchestration of heterogeneous asynchronous multi-source evidence, with operator-retunable objectives and without retraining cycles) is directly applicable to commercial satellite-constellation tasking and cross-mission sensor-resource management; maritime-domain-awareness products for port authorities, coast guards, and commercial insurers; critical-infrastructure monitoring across electricity, water, transportation, and pipeline sectors; supply-chain security and provenance verification; public-health biosurveillance; sensor-resource management for commercial autonomous fleets; and enterprise security-operations-center augmentation in which heterogeneous multi-modal indicators must be correlated under cost, latency, and regulatory constraints.

References

1. U.S. Department of Defense. “Summary of the Joint All-Domain Command and Control Strategy.” Department of Defense, 2022.

2. Gabrielle Kaili-May Liu, Areeb Gani, Jacqueline Lu, Jordan Thomas, Mark Steyvers, Arman Cohan. “Metacognition in LLMs: Foundations, Progress, and Opportunities.” ArXiv preprint arXiv:2607.11881v1, 2026.

3. Hua Wei, Paulo Shakarian, Christian Lebiere, Bruce Draper, Nikhil Krishnaswamy, Sergei Nirenburg. “Metacognitive AI: Framework and the Case for a Neurosymbolic Approach.” arXiv:2406.12147v1, 2024.

4. Aditya Taparia, Noel Ngu, Mario Leiva, Joshua Shay Kricheli, John Corcoran, Nathaniel D. Bastian, Gerardo Simari, Paulo Shakarian, and Ransalu Senanayake. “VLC Fusion: Vision-Language Conditioned Sensor Fusion for Robust Object Detection.” arXiv preprint arXiv:2505.12715, 2025.

5. Shahriar Rifat, AQM Sazzad Sayyed, Nathaniel D. Bastian, and Francesco Restuccia. “Metacognitive Artificial Intelligence in Vision Foundation Models: Research Challenges.” IEEE Intelligent Systems, 2026.

Keywords

Metacognitive AI; Causal Reasoning; Compositional Reasoning; Graph-Based Reasoning; Neuro-Symbolic AI; Hyperdimensional Computing; Human-Machine Formation; Multi-INT Orchestration; Persistent ISR; Value of Information; Explainable AI; Auton

Frequently asked questions

What is DPA27BZ02-DV002?

It is a DARPA SBIR topic in the DoW SBIR 2027 BAA titled Autonomous Metacognitive Collection Orchestration for Multi-Domain Intelligence. It seeks an AI substrate that orchestrates multi-modal intelligence collection and produces traceable, causally informed recommendations.

When is DPA27BZ02-DV002 due?

Proposals are due January 6, 2027, 12:00 PM ET. Pre-release began November 4, 2026, and direct contact with the topic authors (TPOCs) is allowed only until the topic opens on December 2, 2026, 8:00 AM ET.

Is this Phase I or Direct-to-Phase II?

Direct-to-Phase II only. Phase II is described as a 24-month period that takes the capability from TRL 4 at award to TRL 6, staged from at least two modalities in Months 1-12 to at least four in Months 12-24. Note that the Month 21 and Month 24 milestones are listed under an 'Option Milestones' heading, so the base/option split is worth clarifying with the topic authors before December 2, 2026, 8:00 AM ET.

What capstone scenarios must the Phase II prototype handle?

At least three scenarios agreed with the Program Manager, drawn from multi-modal collection under emissions control, cross-INT corroboration against a commander-specified hold-fire confidence threshold, and denial-and-deception resilience with metacognitive adaptation. At least one must involve active adversarial denial or deception.

Book a Free Fit Call