IoT and connected devices is no longer a connectivity story but an edge-compute, energy-autonomy and sensing-reuse story, where the frontier work is squeezing learning onto microcontrollers at nanoamp idle currents and extracting radar-like sensing from radios and fibre that are already deployed.
Summary
“IoT and connected devices” covers the population of low-cost, power-constrained, network-attached endpoints: wireless sensor nodes, wearables, smart-building sensors, tags, and the gateways and edge servers behind them. The category is old in technology terms. One of the sources notes the Internet of Things was introduced almost two decades ago, and that the wearable market has grown continuously over that period. What is moving now is not the idea of connecting a sensor, but four specific engineering constraints that decide whether a deployment is economic.
The first is energy. A node that must be visited to change a battery is usually not worth deploying. Recent work attacks the quiescent current that dominates energy budgets in light-starved photovoltaic installations, using a hardware-orchestrated power manager that fully power-gates the microcontroller and peripherals and wakes it from an ultra-low-power PMIC, RTC and latch, reaching 452 nA idle drain. The second is on-device intelligence. Because backhaul and cloud inference cost power, bandwidth and privacy, the work is on fitting learning, not just inference, into milliwatt budgets: neural architecture search that runs on the embedded device itself under 512 MB of RAM to generate tiny CNNs for low-end MCUs, an accelerator supporting end-to-end on-chip few-shot and continual learning on sequential data at 0.5% area overhead over the inference logic, and a fabricated 65 nm hyperdimensional encoder that uses transistor process variation as its entropy source and reports 7.13 nJ per encoding with 93.2% accuracy on EMG and 96.1% on UCI-HAR.
The third constraint is sensing cost, and this is where the most interesting shift sits. Rather than deploying new sensors, several lines of work reuse infrastructure that already exists. Passive bistatic integrated sensing and communication extracts Doppler-based, radar-like target separation from unmodified LoRa uplinks, validated with two USRP B210 software-defined radios and without interfering with the live network. Passive backscatter tags act as virtual anchors for non-line-of-sight indoor positioning, achieving an aggregated median error of 0.505 m in office-corridor measurements with four devices at 866 MHz. Installed fibre-optic plant is repurposed as a dense seismic array through distributed acoustic sensing, with a graph-theoretic analysis showing a percolation transition at 51.6% coverage above which a city is effectively fully covered for statistical traffic monitoring.
The fourth constraint is heterogeneity and operations. Edge fleets span low-power microcontrollers to accelerator-equipped nodes, which makes deploying AI pipelines across them a scheduling problem rather than a packaging problem, and drives orchestration frameworks that separate allocation decisions from execution. New device substrates carry their own operational tax: flexible IGZO thin-film electronics lack packaging and pins, making conventional automatic test equipment impractical and forcing built-in self-test with 93% single-defect coverage at 3% power overhead. The parameters that decide the field are therefore energy per inference, idle current, sensing accuracy achievable without new hardware, and the cost of managing a heterogeneous fleet.
Evidence base
- 15 June 2026: radar-like Doppler sensing shown feasible using unmodified LoRa signals in a purely passive bistatic configuration, validated with two USRP B210 software-defined radios and without interfering with network operation.
- 16 June 2026: hardware-orchestrated dynamic power management reaches 452 nA quiescent drain by fully power-gating the MCU and peripherals, targeting photovoltaic sensor nodes in low light.
- 16 June 2026: fabricated 65 nm privacy-preserving neuromorphic encoder measures 7.13 nJ per encoding, 76.44 nJ per prediction and 357.32 nJ per training update, with 93.2% EMG and 96.1% UCI-HAR accuracy and federated learning support.
- 16 June 2026: distributed acoustic sensing coverage analysis identifies useful monitoring applications below 10% fibre coverage and a percolation transition at 51.6% coverage above which a city is effectively fully covered.
- 17 June 2026: four passive backscatter devices at 866 MHz deliver 0.505 m aggregated median indoor non-line-of-sight positioning error in office-corridor measurements, without phase synchronisation or transmit-power calibration.
- 20 July 2026: fixed-point DNN denoising on an AMD-Xilinx Kria KV260 achieves 9.7 ms first-sample latency, meeting the 10 ms hearing-aid clinical threshold, while speech separation remains at 16.0 ms.
- 22 July 2026: built-in self-test for flexible IGZO-TFT mixed-signal circuits reports 1100x area reduction and 5600x lower power than prior designs, with 93% single-defect coverage at 3% power overhead, because conventional ATE is impractical for this substrate.
Open questions
- Can passive LoRa ISAC hold Doppler-based target separation in a live multi-node network with uncontrolled traffic, rather than the controlled bistatic setup validated with two USRP B210 radios?
- Is the environmental-change false-positive problem in sub-THz presence detection, where slight furniture movement produces channel variation comparable to human presence, solvable by calibration or does it require a fundamentally different detection metric?
- Does on-chip DAS actually become available at the cost and range assumed by the coverage-threshold analysis, since the entire 51.6% percolation argument is conditional on that assumption?
- Can the data-movement bottleneck identified on embedded FPGA speech enhancement be closed enough to bring speech separation from 16.0 ms under the 10 ms clinical threshold, and does the same bottleneck cap other real-time edge inference workloads?
Assessment drafted 2026-08-31 from up to 18 KB sources using the technology-scorecard framework; scores are a draft read pending review.