A bedroom at three in the morning is an awkward place to put a camera. A road disappearing into fog is a difficult place to rely on one. Above a city, the urgent question might be whether a moving speck is a bird or a drone. Radar has something to offer in each setting. AI is helping it give more useful answers.
Calling this a “second life” needs care. Radar never retired: aviation, weather forecasting and defence kept it busy. Its newer opportunity lies in interpreting the world at a scale and level of detail that more applications can use. Neural networks learn patterns in reflections, motion and time, helping systems recognise activities, classify objects and estimate what happens next. They build on decades of signal processing.1
01The echo is becoming an answer
This revival has several parents. More capable radar chips and antenna designs improve the measurements; AI makes more of those measurements useful. NXP’s 2023 collaboration with Zendar, for example, pursues higher resolution by combining multiple radars into a larger effective aperture. Hardware still matters. An algorithm cannot negotiate with missing information.2
The research is becoming substantial. A 2026 review analysed 85 peer-reviewed papers from 2018–2025 on deep learning for human activity recognition using FMCW radar, a common compact radar architecture. That is one specialised branch, rather than a count of the whole field.3 Capital is following specific opportunities too: MatrixSpace announced a $20 million Series B in October 2025 for its AI-enhanced radar business.4 These are signs of activity; neither establishes that AI alone caused a market revival.
Papers in one review of deep learning for FMCW radar activity recognition. [3]
One disclosed MatrixSpace funding round. A company signal, not a market total. [4]
02A useful kind of discretion
In care, the attraction is immediate: detect a possible fall without installing a video feed or asking someone to remember a wearable. Vayyar sells radar-based fall alerts and AI insights without cameras or sound recording; its sensors were included in Redbridge’s care-technology pilot in 2025.5, 6 The practical gain is a chance to make help less dependent on someone pressing a button.
Sleep research shows how much meaning can emerge from a small signal. A 2025 Nature Communications study adapted a deep-learning model to radar-derived breathing signals, reaching 75.81% accuracy in four-stage sleep classification on its radar dataset.7 That is a research result with errors, and a reminder that discretion is not anonymity. A sensor that reveals sleep or daily routines still produces sensitive information. The privacy opportunity depends on consent, limited retention and control over what leaves the device.
03When the picture gets worse
Radar’s independence from visible light, and its relative robustness in poor weather, become more valuable as machines take on consequential tasks.1 In 2025, Arbe announced work with NVIDIA on AI-driven mapping of drivable space from radar. The aim moves beyond tracking a vehicle ahead to understanding where a vehicle could go.8
At another scale, weather radar supplies the observations for NowcastNet, which combines physical constraints with deep learning to forecast precipitation up to three hours ahead. In its 2023 study, 62 meteorologists ranked it first in 71% of evaluated cases against competing methods.9 Better short-term warnings could give emergency teams time to act. That evaluation measures forecast preference, however, not lives saved.
Defence makes the classification problem especially sharp. A bird mistaken for a drone wastes attention; a missed drone can be catastrophic. Robin Radar uses deep neural networks with micro-Doppler signatures, the small motion patterns of wings or rotors. In 2025, it reported 80% fewer false alarms with a new model at a specified operating setting.10 Demand is concrete: the Dutch Ministry of Defence announced an order for 51 mobile Robin radars for Ukraine in August 2024.11
04What deserves to be trusted?
The next test is operational. Reflections can be ambiguous, training data incomplete and unfamiliar environments unforgiving.1, 3 Buyers should ask about missed events, false alarms per hour, performance on unfamiliar sites and behaviour when confidence collapses. A care alert needs a responder. A defence classifier needs accountable human decisions. A convincing demonstration settles neither question.
For CHIRP, the compelling dual-use opportunity is the shared demand for dependable perception: protecting people in private spaces, keeping infrastructure working and understanding threats. The applications need different safeguards, even when they share engineering.
Radar’s second life will be earned wherever AI turns an echo into a decision people can trust. The most interesting question is how much understanding we can gain while collecting only what the task requires.
05Sources and evidence notes
Sources checked on 28 September 2026. The argument is CHIRP’s editorial interpretation of the evidence. Research findings, product announcements and procurement records are identified separately below.
How to read the numbers
Publication activity. The 85-paper figure is the inclusion count reported by Dang and colleagues for one subfield across 2018–2025. It is neither an annual time series nor a worldwide count of “AI + radar”, and it does not establish growth by itself. A separate 2024 review by Sénica, Marques and Figueiredo documents growing IEEE Xplore search activity for “Artificial Intelligence Radar” across 2015–2022 (Figure 4). Its historical search is not extrapolated to 2026 here.
Entrepreneurship. MatrixSpace’s $20 million Series B is a company-disclosed round, not sector-wide investment or revenue. It is evidence of financing for a radar-and-AI venture. A reliable annual census of newly founded radar startups was not established in this research, so no global startup count or growth rate is claimed. MatrixSpace dates its founding to late 2019 in its 2022 launch announcement.
Sleep research. The 75.81% figure applies specifically to four-stage classification on the ClinRadar dataset. The paper’s larger training corpus contains respiratory signals from other sensing methods; it is not a 15,785-night radar dataset. The result does not establish a universal accuracy or a replacement for clinical assessment.
Weather. NowcastNet’s 71% is the reported share of evaluated cases in which experts preferred its forecast over the comparison methods. It is not a general “71% forecast accuracy”, and the study does not measure emergency-response outcomes. Weather radar and compact room or vehicle radars are different systems; their common thread here is learned interpretation of radar observations.
Defence and Figure 01. The ministry confirms a purchase of 51 mobile radars intended for Ukraine, rather than an independently verified delivery or performance result. Robin’s 2025 comparison concerns different neural-network versions, not AI versus a purely conventional baseline. It reports an 80% false-alarm reduction when thresholds preserve similar classification ranges for larger drones, with doubled ranges for smaller drones. The chart simply indexes the previous level to 100 and calculates 100 × (1 − 0.80) = 20; neither number is a measured count of alarms per hour.
Causation. These sources support a widening range of applications and specific examples of progress. They do not isolate AI’s contribution to overall radar market growth from advances in hardware, available computation, regulation or changes in demand.
- Srivastav & Mandal. Radars for Autonomous Driving: A Review of Deep Learning Methods and ChallengesResearch review · IEEE Access · 2023
Radar capabilities, learning methods, uncertainty, clutter and dataset limitations. Published version.
- NXP and Zendar accelerate high-resolution radar developmentCompany announcement · 2 November 2023
Distributed aperture radar and semiconductor integration illustrate the complementary contribution of hardware.
- Dang et al. Deep Learning-based Human Activity Recognition with FMCW Radar: A ReviewResearch review · IEEE Sensors Journal · 2026
Author-hosted abstract reports 85 peer-reviewed publications from 2018–2025 and identifies remaining robustness and evaluation gaps. Publisher record.
- MatrixSpace completes $20M Series B fundingCompany announcement · 15 October 2025
Disclosed investment in an AI-enhanced radar company. Used as an individual financing example.
- Vayyar Care AIManufacturer product information · Accessed September 2026
Camera-free and audio-free monitoring, fall alerts and AI insights. Its technology description links radar data with machine-learning applications. Capability claims are the manufacturer’s.
- Care City. Redbridge Care Tech Pilot: A Year of Innovation in Falls PreventionImplementation partner report · 4 September 2025
Confirms Vayyar sensors among the technologies being piloted. No programme-wide outcome is attributed to radar alone.
- Zhuang et al. Advancing sleep health equity through deep learning on large-scale nocturnal respiratory signalsPeer-reviewed study · Nature Communications · 22 October 2025
Results section and Table 2 distinguish radar-specific sleep classification from respiratory-belt results.
- Arbe collaborates with NVIDIA on radar-based free-space mappingCompany announcement · 6 January 2025
AI processing of high-resolution radar data for vehicle perception. An announced development and demonstration, not proof of fleet-wide safety.
- Zhang et al. Skilful nowcasting of extreme precipitation with NowcastNetPeer-reviewed study · Nature · 5 July 2023
Radar-based forecasts with lead times up to three hours; expert evaluation by 62 meteorologists. DOI.
- Robin Radar. Deep neural networks in the IRIS radarManufacturer technical account · 3 February 2025
Source for micro-Doppler classification and the conditional 80% false-alarm reduction used in Figure 01.
- Dutch Ministry of Defence. Mobile drone-detection radars for UkraineGovernment procurement announcement · 21 August 2024
Purchase of 51 Robin mobile radars. English ministry summary, 28 August 2024.