Part of the DoW SBIR/STTR FY27 Release 2 topic list. Browse all 83 topics by industry or service.
Topic DTR27BZ02-NV003 from the DoW SBIR 2027 BAA, Release 2. Pre-release now; proposals due January 6, 2027, 12:00 PM ET.
Quick answer
The Defense Threat Reduction Agency (DTRA) wants an explainable machine-learning decision engine that tells an autonomous platform which geophysical sensor to run, inhibit, or escalate to while searching for subsurface hazards, including potential WMD devices or components. This is an SBIR topic that starts with a Phase I feasibility effort. Proposals are due January 6, 2027, 12:00 PM ET.
Solicitation at a glance
- Topic number: DTR27BZ02-NV003
- Official title: Dynamic Sensor-Tasking and Routing System for Subsurface Hazard Detection
- Agency: Defense Threat Reduction Agency
- Program: SBIR (DoW SBIR 2027 BAA, Release 2)
- Award type: Phase I (see topic text for Phase II)
- Award amount: Up to $250,000 (7 months, plus up to $6,500 TABA). Source: DTRA Release 2 instructions in DSIP.
- CMMC level: Level 1
- ITAR: No
- Technology areas: Nuclear, Sensors
- 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 DTR27BZ02-NV003)
BW&CO's take
What the Defense Threat Reduction Agency wants
Today's unmanned platforms run every subsurface sensor continuously, which floods tactical networks, slows the mission, and keeps collecting useless data when soil conditions degrade a sensor, such as ground-penetrating radar in wet, conductive clay. DTRA wants learned decision policies (reinforcement learning, contextual bandits, or POMDPs) that use sensor-inferred soil physics, mission constraints, and live sensor performance to queue, activate, or inhibit sensors, decide when an anomaly needs a higher-fidelity second sensor, and adjust routing. Phase I builds a software emulator or conceptual model and ends with a demonstration showing less survey time and data overhead than static multi-sensor collection. Phase II delivers edge software that runs on embedded processors without cloud reachback, tested with real sensor data.
Who is a good fit
- Hands-on experience with reinforcement learning, contextual bandits, or POMDPs for sequential decisions under uncertainty
- Working knowledge of EMI, GPR, and magnetometer phenomenology, including how electrical conductivity and dielectric permittivity affect each sensor
- Real-time ML inference on embedded processors with no cloud or high-performance computing support
- Access to real geophysical survey data, for example from munitions response (UXO) or subsurface utility mapping work
Watch-outs
- Phase I must conclude with a demonstration that proves a reduction in survey time and data overhead versus static multi-sensor acquisition, so plan for a working demonstration, not only a paper study.
- The objective calls for an explainable architecture; a black-box policy with no account of why it gated a sensor is a weak fit.
- Phase II demonstrations use real physical sensor data, so show where that data will come from.
- The topic lists CMMC Level 1. Direct contact with the topic authors is allowed only during pre-release, which ends when the topic opens December 2, 2026, 8:00 AM ET.
Think your company fits DTR27BZ02-NV003? Talk with BW&CO before the January 6 deadline.
Plan Your DTR27BZ02-NV003 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
Dynamic Sensor-Tasking and Routing System for Subsurface Hazard Detection
Objective
Develop an explainable, adaptive machine learning architecture using Super Intelligence (SI) that dynamically schedules, sequences, and gates geophysical sensor modalities based on real-time environmental physics, learned sensor operational envelopes, and predictive target response models. The system will enable autonomous platforms to perform in-stride characterization and clearance of subsurface hazards, including potential WMD devices or components, in high-tempo, contested tactical environments.
Description
Current field-deployable capabilities for WMD device and component defeat, route reconnaissance, and sensitive site exploitation have made significant strides in rapid surface-level mapping and visual threat triage using unmanned aerial and ground platforms. However, these operational systems encounter a critical bottleneck when confronting subsurface hazards. Visual, optical, and topographic sensors are fundamentally blind to threats that are partially buried, completely subsurface, or obscured beneath soil, blast debris, or disturbed earth. When subsurface geophysical instruments are integrated onto these unmanned platforms to address this blind spot, current field architecture suffers from a severe "collect everything" bottleneck. Lacking localized intelligence, platforms typically execute static, pre-programmed survey scripts which force sensors to operate continuously and indiscriminately across an entire area. This floods constrained tactical communication networks with massive volumes of raw sensor data, overwhelms field computing devices, drastically slows operational tempo, and results in sensor operational blindness when local soil physics degrade or invalidate specific sensor modalities. When conducting standoff reconnaissance of suspected hazard areas—which may contain buried chemical or biological munitions, radiological dispersal device (RDD) components, bulk precursor storage vessels, or associated explosive triggers and conventional ordnance—these bottlenecks force operators into dangerous, prolonged exposure on target. Furthermore, relying on rigid, static classical logic to manage these complex sensor suites in real time introduces a critical limitation, as traditional rule-based programs cannot adapt to highly dynamic, uncharacterized environments where the boundary conditions are not known beforehand.
A proactive system is required to answer:
"Given this specific environment and mission profile, which sensor should be deployed first, what detection threshold triggers the next modality, and which sensors may be inhibited due to the environment or other physical boundary conditions?"
This topic seeks an adaptive, machine-learning-driven sensor-tasking and routing decision engine capable of scheduling sensor deployment and routing platform payloads in real time using probabilistic, data-driven optimization methods. The primary need is an intelligent edge capability that continuously monitors changing site conditions and dynamically queues, activates, or inhibits physical sensor payloads on an autonomous platform. Rather than streaming raw data or running all sensors simultaneously, the system must govern sensor execution through learned decision policies (utilizing techniques such as reinforcement learning, contextual bandits, or partially observable Markov decision processes) that optimize performance based on sensor-inferred environmental physics (such as electrical conductivity and dielectric permittivity), immediate mission constraints (such as required advance speed, standoff distance requirements, and target priority), and real-time sensor performance signatures. When an initial screening modality detects an anomaly, the engine must evaluate the environmental context to dynamically determine if a secondary, higher-fidelity sensor is required to verify and classify the threat, or if the platform should continue its route. By automating this sensor-queuing and routing logic at the edge through explainable and efficient AI, the system will overcome the subsurface bottlenecks of current field suites, eliminate tactical network saturation, minimize time on target, and ensure reliable detection and defeat of subsurface WMD devices and associated hazards across diverse, uncharacterized operational environments.
State of Technology: Currently deployed military and commercial technologies have reached a high level of maturity in distinct, stove-piped domains, but they do not integrate these capabilities into a single, closed-loop, in-stride subsurface decision engine. The current technical landscape is characterized by the following:
- Standoff Aerial and Surface Visual Mapping Systems: Existing fielded systems utilize unmanned aerial vehicles (UAVs) and ground platforms to perform rapid photogrammetric, optical (EO/IR), and LiDAR mapping of surface terrains. These systems feature mature edge-computing architectures that run real-time automated target recognition (ATR) to classify surface munitions, craters, and obstacles on tactical hand-held tablets in under 30 minutes. However, these systems lack subsurface sensing capability, limiting their ability to detect and characterize hazards that are partially buried, covered by debris, or below the ground surface.
- Geophysical Post-Processing Classification Suites: For subsurface detection, advanced electromagnetic induction (EMI) and ground-penetrating radar (GPR) systems are highly mature and accredited for DoW munitions response. These systems utilize sophisticated physics-based inversion models to extract target polarizability tensors and classify anomalies against standard munitions libraries. However, these software suites are strictly designed for post-processing desktop environments. They are not integrated into real-time, in-stride robotic platforms, and they cannot make autonomous decisions in the field regarding active sensor scheduling.
- Uncoordinated "Collect-Everything" Sensor Payloads: Current multi-sensor ground platforms operate in a fixed, concurrent acquisition mode where all integrated sensors (such as magnetometers, GPR, and EMI) run continuously regardless of whether the local terrain and environment is physically compatible with the sensor. For example, when a platform transitions into wet, highly conductive clay, a ground-penetrating radar (GPR) antenna suffers from severe signal attenuation, rendering the collected data ineffective for its intended detection purpose. Current systems lack the real-time feedback loops to recognize these environmental changes and limitations then automatically adjust, disable, or redirect sensor payloads. This lack of dynamic management creates a major operational inefficiency, resulting in the collection of highly degraded data that requires excessive post-processing, while unnecessarily consuming onboard battery power, communication bandwidth, and valuable mission time.
- Static Autonomous Path Planning: Existing autonomous ground and aerial navigation algorithms plan paths based on surface traversability, geometric obstacle avoidance, and pre-mapped terrain features. They do not ingest real-time subsurface hazard maps or dynamically adjust their routing, payload dwell times, or sensor modalities based on changing subsurface soil electromagnetic properties or detected anomalies.
- Adaptive AI/ML for Sequential Decision-Making and Path Planning: In adjacent fields such as autonomous warehouse robotics, dynamic network routing, and space domain awareness, machine learning is highly mature for solving complex, real-time sensor scheduling and path-planning challenges. State-of-the-art approaches utilize Deep Reinforcement Learning (DRL) and Contextual Multi-Armed Bandits to train autonomous agents that make optimal sequential decisions under extreme uncertainty. These AI models observe dynamic environmental contexts (e.g., localized obstacle density or changing wireless signal-to-noise ratios), weigh the trade-offs between exploring unknown areas and exploiting known paths, and automatically adjust sensor power cycles or update the platform’s route on-the-fly. However, these advanced decision-making frameworks have not yet been integrated with or trained on subsurface geophysical physics, leaving a critical gap where current autonomous AI agents remain incapable of managing multi-modal subsurface sensor suites.
Phase I
Demonstrate the technical and operational feasibility of an adaptive, machine-learning-Super Intelligence (SI) driven decision engine for dynamic sensor-tasking and routing. Efforts will focus on identifying the theoretical machine learning models, policy optimization frameworks, and reward structures (e.g., reinforcement learning, contextual bandits, or other decision-making algorithms) governing a representative sensor suite, which may include multiple sensor phenomenology’s to identify representative subsurface and surface hazards, including potential WMD devices or components.
The investigator will develop an Super Intelligence (SI)-based software emulator or conceptual model to evaluate the feasibility of the learned sensor-gating policies and decision rules under varying environmental and mission constraints. Phase I will conclude with a demonstration showing that the adaptive, machine-learning-driven system can successfully execute sensor-gating, anomaly escalation, or vehicle routing to prove a reduction in survey time and data overhead compared to static multi-sensor acquisition.
Phase II
Transition the theoretical machine learning models, policy optimization frameworks, and adaptive sensor-gating logic developed in Phase I into a functional software prototype of the decision engine. The investigator will design and program an edge-deployable software package optimized to execute real-time model inference on embedded processors without requiring cloud reachback or high-performance computing resources during operations.
The prototype design and engineering effort will focus on converting Phase I's learned decision models and policy algorithms into robust, real-time software capable of ingesting in-stride environmental feeds and continuously calculating the optimal sensor-tasking sequence at the tactical edge. The investigator will conduct demonstrations using real physical sensor data to verify that the machine-learning-driven decision engine successfully executes the correct sensor-sequencing commands (activating, disabling, or swapping sensor modalities based on physical environmental transitions) and adapts simulated platform routing based on learned, unclassified target-detection thresholds.
Phase III Dual Use Applications
Phase III will scale the testing and validation of the machine-learning-driven software prototype developed in Phase II through larger-scale field demonstrations and expanded evaluations across highly diverse, uncharacterized environments. This phase will focus on scaling the system's capabilities by validating the generalization of the trained decision policies against significantly larger, multi-site real-world environmental datasets and conducting testing across larger geographical search areas to ensure the model adapts to novel terrain without requiring real-time retraining. Testing and evaluations should focus on demonstrating the software's execution on representative unmanned ground vehicle (UGV) and unmanned aerial system (UAS) platforms configured with the sensor payloads identified in the previous phases.
The expanded demonstrations will evaluate the SI decision engine's ability to repeatably execute sensor-gating commands and adapt platform routing when encountering a wider array of subsurface hazards and simulated WMD-relevant components under diverse, live-field terrain and soil conditions. This expanded validation will provide the necessary performance data to support future technology transition pathways into DTRA C-WMD technology suites, Joint Service EOD robotic systems, and automated route-reconnaissance and breaching suites to support future engineering requirements. For commercial dual-use, the scaled software will transition to geophysics and environmental remediation contractors conducting UXO clean-up at Formerly Used Defense Sites (FUDS), as well as subsurface utility engineering (SUE) platforms to enable more efficient, rapid subsurface mapping in complex civil environments.
References
F. Samadzadegan, , and S. M. Mousavi, “Unmanned Aerial Remote Sensing for Unexploded Ordnance Detection: Concepts, Methods, Current Status and Trends,” Sens Imaging, vol. 26, no. 1, p. 137, Oct. 2025, doi: 10.1007/s11220-025-00668-5.
D. R. Glaser, “Standoff High-Frequency Electromagnetic Induction Response of Unsaturated Sands: A Tank-Scale Feasibility Study,” Journal of Environmental and Engineering Geophysics, vol. 27, no. 1, pp. 45–51, Mar. 2022, doi: 10.32389/JEEG21-030.
M. D. Richardson and M. T. Tuley, “SERDP/ESTCP Munitions Response Program: A Summary of Progress and a Way Forward for Underwater Remediation of Unexploded Ordnance (UXO),” in OCEANS 2022, Hampton Roads, Hampton Roads, VA, USA: IEEE, Oct. 2022, pp. 1–6.
doi: 10.2113/JEEG17.2.53.
doi: 10.3390/e24091315.
doi: 10.1117/12.2663970.
doi: 10.3390/s25196076.
doi: 10.3390/s25196076.
Keywords
Unmanned Aerial System (UAS): Explosive Ordnance Disposal (EOD): Remote Render Safe: Precision Aiming: Operational Standoff: Disruptor Payload: CBRN Environments: Remote Command and Control: Trusted Autonomy: Small-Unit UAS
Frequently asked questions
What is DTR27BZ02-NV003?
It is a DTRA SBIR topic in the DoW SBIR 2027 BAA seeking an adaptive, explainable machine-learning engine that dynamically schedules, sequences, and gates geophysical sensors and routes autonomous platforms to detect subsurface hazards, including potential WMD devices or components.
When is DTR27BZ02-NV003 due?
Proposals are due January 6, 2027, 12:00 PM ET. Pre-release began November 4, 2026, and direct contact with the topic authors is allowed only until the topic opens on December 2, 2026, 8:00 AM ET.
Is this a Phase I or Direct-to-Phase II topic?
The topic describes a Phase I feasibility effort (an emulator or conceptual model plus a demonstration), followed by a Phase II software prototype. It does not describe a Direct-to-Phase II option.
What decisions does the engine have to make?
It must answer which sensor to deploy first, what detection threshold triggers the next modality, and which sensors to inhibit because of environmental conditions. When a screening sensor detects an anomaly, the engine decides whether a higher-fidelity sensor is needed to verify it or whether the platform should continue its route.