Raspberry Pi: Versatile Embedded Platform

• Raspberry Pi is a versatile, low-cost, low-power single-board computer offering rich peripheral access and a full Linux environment.
• It supports diverse applications including biosignal acquisition, edge AI, real-time control, and cluster computing with optimized software implementations.
• Research demonstrates its effectiveness in enabling machine learning, distributed sensor networks, and innovative educational and experimental setups.

Raspberry Pi (RPi) denotes a class of low-cost, low-power single-board computers that, in the research literature surveyed here, appears less as a single-purpose hobbyist board than as a reconfigurable embedded computing substrate. Across the cited studies, RPi Zero, Zero W, Raspberry Pi 3 Model B+, Raspberry Pi 4 Model B, and Raspberry Pi 5 Model B serve as biosignal hosts, edge-AI targets, laboratory controllers, cluster nodes, vehicular communication stations, RFID data loggers, portable SDR controllers, and autonomous star sensors, with Linux software stacks coupled to GPIO, SPI, I²C, CSI, UART, USB, PCIe, Ethernet, and Wi‑Fi peripherals (Youngblood, 2020, Margapuri, 2020, Ju et al., 2022, Farina et al., 19 Mar 2025).

1. Platform profile and hardware envelope

The papers collectively characterize Raspberry Pi as a low-cost, low-power embedded platform whose practical envelope is defined by modest CPU throughput, limited RAM, and unusually rich peripheral I/O. In the cited deployments, this envelope ranges from the Raspberry Pi Zero W used in a battery-powered RFID birdfeeder, to Raspberry Pi 3 Model B+ systems used for motion detection, secure home automation, and browser-accessible MPI clusters, to Raspberry Pi 4 Model B nodes used for CNN inference, Bayesian optimization, ImageNet-scale linear training on precomputed features, and distributed containerized edge computing, and finally to Raspberry Pi 5 Model B systems used as full ETSI C-ITS stations with PCIe-attached 5.9 GHz radios (Youngblood, 2020, Imran et al., 2020, Gupta et al., 2021, Ju et al., 2022, Do, 2022, Mahmud et al., 2021, Farina et al., 19 Mar 2025).

Raspberry Pi model	Representative role	Source
Raspberry Pi Zero W	RFID-equipped birdfeeder, Wi‑Fi logger, cloud client	(Youngblood, 2020)
Raspberry Pi 3 Model B+	Motion detector, SMTP-based home automation node, web-facing HPC front end	(Margapuri, 2020, Gupta et al., 2021, Imran et al., 2020)
Raspberry Pi 4 Model B	CNN target platform, auto-aligner controller, ImageNet incremental local SGD node, Con-Pi edge/fog node	(Ju et al., 2022, Mathew et al., 2020, Do, 2022, Mahmud et al., 2021)
Raspberry Pi 5 Model B	OBU/RSU running OScar with 5.9 GHz ITS-G5 radio	(Farina et al., 19 Mar 2025)
Raspberry Pi Zero	Autonomous star sensor computer	(P et al., 2022)

At the hardware level, the studies repeatedly stress constraints that directly shape algorithm and system design. Raspberry Pi 4 Model B is evaluated with 4 GB RAM and a quad-core Cortex-A72 at 1.5 GHz, with inference executed on the ARM CPU rather than a CUDA-capable GPU; Raspberry Pi Zero provides only a single-core 1 GHz BCM2835 with 512 MB DDR2; Raspberry Pi 5 contributes PCIe 2.0 expansion, 8 GB LPDDR4, and enough compute to host a full OScar stack, but requires a power supply able to deliver at least 25 W at 5 V when combined with a PCIe radio (Ameen et al., 2023, P et al., 2022, Farina et al., 19 Mar 2025). This hardware profile explains why the papers emphasize direct peripheral coupling, low-overhead software paths, and careful balancing of compute, memory traffic, and energy.

2. Interfaces, operating systems, and implementation patterns

A consistent pattern across the literature is that Raspberry Pi is valuable not only because it is inexpensive, but because it exposes a Linux environment and a broad peripheral surface simultaneously. The cited systems use Raspberry Pi OS or Raspbian, standard Linux networking, and common language/runtime combinations such as C, Python, Java, PyTorch, TensorFlow Lite, Flask, and Ansible. Peripheral access is realized through SPI for ADS1299 EEG acquisition and MCP3008 ADCs, I²C for ADS1115 photodiode readout, CSI for camera modules, UART for MIDI and telemetry, USB for SDR dongles and GNSS receivers, PCIe on Pi 5 for miniPCIe Wi‑Fi cards, and GPIO for relays, LEDs, steppers, and multiplexers (Rakhmatulin et al., 2022, Mathew et al., 2020, Barbancho et al., 2024, Farina et al., 19 Mar 2025, Choudhury et al., 2021).

Several papers make the software split explicit. PiEEG uses a low-level data acquisition library in C, compiled as a static library, and a high-level Python layer for filtering, FFT, feature extraction, and GPIO control; the authors justify this split by stating that “when a Python file is corrected, the library does not need to be recompiled” (Rakhmatulin et al., 2022). The auto-aligner places motor control and ADC interaction on the Pi under Python while using M‑LOOP’s Gaussian-process Bayesian optimization through a file-based API (Mathew et al., 2020). LibrePiLogger uses Python drivers in AtmosPyre for Modbus sensors and Ansible to deploy virtual environments, generated logging scripts, and systemd services from a single YAML inventory file (Keim et al., 11 May 2026). Con-Pi uses Flask-based REST services, Docker containers, and Crontab startup on Raspbian Stretch or Buster to distribute control-plane and sensor-plane functions across multiple RPis (Mahmud et al., 2021).

This suggests a recurrent RPi implementation idiom: hardware-near operations remain in C, dedicated MCU firmware, or tightly scoped services, while orchestration, policy, and experimentation move into Python, REST, or declarative deployment tooling. The same idiom reappears in high-speed SPI biosignal capture, RS‑485 data logging, optical alignment, distributed edge computing, and vehicular networking (Otte et al., 2018, Rakhmatulin et al., 2022, Keim et al., 11 May 2026, Mahmud et al., 2021).

3. Measurement, sensing, and data acquisition

One of the most thoroughly instrumented uses of Raspberry Pi in the corpus is biosignal acquisition. PiEEG is a shield built around the ADS1299 and mounted on the Raspberry Pi 40-pin header, exposing 8 channels for EEG electrodes, 1 reference channel, and 1 bias channel, with 24-bit resolution and SPI transfer rates from 250 SPS to 16 kSPS. The Raspberry Pi configures and reads ADS1299 over SPI, receives data “directly to the processor,” applies a 1–30 Hz band-pass filter, computes an FFT,

$X[k] = \sum_{n=0}^{N-1} x[n] e^{-j 2\pi kn / N},$

and detects blink-related peaks by thresholding amplitude within user-selected bands such as 3–7 Hz or 1–3 Hz. In the blink-control experiments, the system correctly detects and responds in 8 out of 10 cases, with an overall real-time delay of about 1 s in the current implementation. A related PiEEG paper generalizes the same shield to EEG, ECG, EMG, and EOG, with explicit battery-only powering for safety and a C-plus-Python software path on Raspberry Pi OS (Rakhmatulin et al., 2022, Rakhmatulin, 2023).

A second family of sensing systems uses Raspberry Pi as an edge logger rather than a signal-processing front end. The RFID birdfeeder based on Raspberry Pi Zero W couples a perch antenna and a CognIoT 125 kHz RFID reader to the Pi, stores PIT-tag detections with timestamps, and uploads logs via Wi‑Fi using rclone; a 20,000 mAh power bank provides about 4 days of continuous operation (Youngblood, 2020). LibrePiLogger turns Raspberry Pi Zero WH or Raspberry Pi 4 into a Modbus/RS‑485 environmental logger, with hardware costs of 53.82 EUR for the minimal HAT setup and 62.35 EUR for the maximal USB-converter setup, logging timestamped CSV files with JSON metadata; the system has been deployed since spring 2025 for continuous CO$_2$ and $^{222}$Rn monitoring, and “all expected measurements were recorded successfully, with no gaps or corrupted records during system operation” (Keim et al., 11 May 2026). In HPC slow control, the Sensor Aggregation Unit combines a Raspberry Pi, an Arduino, and PoE; the Arduino samples sensors at 1 Hz and forwards data over a serial link, while the Pi integrates the unit into Ganglia or EPICS and even reflashes the Arduino remotely via avrdude (Otte et al., 2018).

Raspberry Pi also appears as a local security and media-acquisition host. In the smart motion detection system, Raspberry Pi 3 Model B+ reads a PIR sensor digital HIGH of approximately 3 V on GPIO, optionally drives an LED, captures either a still image or a 5 s .h264 video from the Raspberry Pi Camera, converts it to .mp4 using GPAC MP4Box, and sends notifications via SMTP. The same work models the system in UPPAAL and verifies reachability, timing bounds, and A[](not deadlock) (Margapuri, 2020). Taken together, these systems present RPi as a practical measurement node for biosignals, RFID telemetry, industrial serial buses, camera-triggered surveillance, and long-running field logging.

4. Edge AI, model optimization, and on-device training

The machine-learning papers in the corpus treat Raspberry Pi not as a generic Linux box but as a specific target architecture whose ARM CPU, memory hierarchy, and thermal budget materially affect model design. In "Efficient Convolutional Neural Networks on Raspberry Pi" (Ju et al., 2022), all experiments are performed on Raspberry Pi 4 Model B with 4 GB RAM, Python 3.9, and PyTorch 1.11.0. The proposed TripleNet uses five Triple-Blocks built from three layer types: a DenseNet-like Conv Layer 1,

$x_n = H_n\left([x_1,\ldots,x_{n-1}]\right),$

an improved harmonic dense Conv Layer 2 with sparse exponentially spaced connectivity,

$x_{n}=\left\{ \begin{array}{ll} H_{n}\left([x_{n-1}]\right), & \text{if } n \% 2 = 1\[4pt] H_{n}\left([x_{0}, \dots, x_{n - 2^i}]\right), & \text{if } n \% 2 = 0 \end{array} \right.,$

and a residual bottleneck Conv Layer 3,

$x_{n+1} = x_n + \mathcal{F}(x_n, w_n).$

On CIFAR‑10, TripleNet-S achieves 4.06 s per 100 images and TripleNet-B achieves the best error, 12.97%; on SVHN, TripleNet-S reaches 3.87 s per 100 images and TripleNet-B the best error, 5.67%. The abstract summarizes that TripleNet shortens inference time per image by 15%, 16%, 17%, 24%, and 30% relative to GhostNet, MobileNet, ThreshNet, EfficientNet, and HarDNet, respectively (Ju et al., 2022).

"Optimizing Deep Learning Models For Raspberry Pi" (Ameen et al., 2023) sharpens that architectural perspective by showing that deployment choices on Pi 4B can dominate nominal sparsity. The paper evaluates Keras, TensorFlow Lite float, TensorFlow Lite quantized, and Arm NN delegate variants on 64-bit Raspberry Pi OS (Debian bullseye) using the Broadcom BCM2711 Cortex‑A72. Unstructured magnitude pruning is tested at sparsity levels from 0.25 to 0.99, but the key empirical conclusion is that pruning does not reduce model size or inference time in their TensorFlow Lite setting because dense kernels still execute dense arithmetic. By contrast, runtime conversion and backend selection can be decisive: on MNIST, TensorFlow Lite reduces mean inference time from approximately 114.26 ms to approximately 3.74 ms, and quantized models to approximately 3.06 ms; on VGG16 for Horses vs Humans, the Arm NN delegate reduces mean inference time from approximately 3475.11 ms to approximately 458.53 ms; yet naive post-training quantization on MobileNetV2 for Cats vs Dogs collapses accuracy to 18.75% (Ameen et al., 2023).

The corpus also includes a rare example of large-scale training on Raspberry Pi. "ImageNet Challenging Classification with the Raspberry Pi" (Do, 2022) trains a multiclass linear classifier on 2,048-dimensional Inception v3 features from the 1,261,405-image ILSVRC 2010 dataset using incremental local SGD. The method partitions the data into $T=8$ blocks, each about 127,000 datapoints, and within each block learns $k=300$ local SGD models. The global objective is standard hinge-loss regularized empirical risk minimization,

$$\min_{w_p} \ \Psi(w_p, [X, Y]) = \frac{\lambda}{2}\|w_p\|^2 + \frac{1}{m}\sum_{i=1}^m \max\{0, 1 - y_i (w_p \cdot x_i)\},$$

with block-level ensembles aggregated by majority vote. On Raspberry Pi 4, this incremental local SGD uses about 2 GB memory, trains in 129.48 minutes, and reaches 75.61% accuracy, outperforming the reported full-SGD and LIBLINEAR baselines that run on a PC with 32 GB RAM (Do, 2022).

5. Real-time control, communication systems, and autonomous devices

Raspberry Pi is repeatedly used as the main compute-and-control element in cyber-physical systems whose peripherals are heterogeneous but whose control path remains local. The Raspberry Pi Auto-aligner uses a Raspberry Pi 4 Model B, GPIO-driven stepper motors, an ADS1115 16-bit ADC on I²C, and the M‑LOOP optimizer in Gaussian-process mode to maximize coupling into a single-mode optical fiber. The parameter vector is

$\mathbf{p} = (p_1, p_2, p_3, p_4),$

with cost

$_2$0

and the platform reaches coupling efficiencies comparable to or slightly better than an experienced human from poor initial alignment, with roughly 20 minutes for 200 evaluations (Mathew et al., 2020).

In vehicular networking, Raspberry Pi 5 Model B becomes a full ETSI C-ITS station when paired with a PCIe HAT, a Mikrotik R11E‑5HND ath9k-based 5.9 GHz radio, and the OScar stack. OScar runs reception, transmission, HMI, Local Dynamic Map maintenance, API, CAN, and GNSS threads on Linux, supports CAM, VAM, CPM, and L0 security, and is validated against EN 302 571. Power spectral density is computed as

$_2$1

and the transmitted spectrum lies below the EN 302 571 mask. The reported platform cost is approximately 200 EUR; CPU load is about 25% for reception only and about 50% when generating CAMs; temperature rises from about 60 °C to about 80 °C during sustained transmission; and line-of-sight range is about 560 m, with reception persisting about 70 m behind a building in one NLOS scenario (Farina et al., 19 Mar 2025).

A distinct navigation application is the low-cost star sensor StarberrySense, where Raspberry Pi Zero executes the entire Lost-In-Space chain: image acquisition through the IMX219 camera, thresholding at

$_2$2

centroiding,

$_2$3

catalogue-based geometric voting, and QUEST attitude estimation. The device achieves 27.18 arcsec pointing-axis accuracy at $_2$4, 38.76 arcmin roll-axis accuracy at $_2$5, 94.69% star-identification success, and about 0.2 Hz quaternion output, with image capture and readout alone taking about 3.6 s on Pi Zero. Average power is below 1.25 W, but the platform is explicitly described as not radiation-hardened and primarily suitable for short-duration LEO missions (P et al., 2022).

Raspberry Pi also anchors lower-cost communication and control stacks. In the portable SDR system, Raspberry Pi Model B acts as baseband processor and control host for an RTL-SDR front end, handles 2.4 MSPS I/Q streams in C, and uses GPIO plus PWM hardware as an FM transmitter; selected reported SNR values include about 34.6 dB for WFM and about 28.6 dB for DSB under the tested settings (Choudhury et al., 2021). In the secure home automation prototype, a Raspberry Pi Model B+ node called Unitor connects to Ethernet, controls relays driving fans, and processes SMTP commands secured by sender filtering, subject filtering, and the Edon80 stream cipher, with the authors noting that 26 GPIO-capable pins allow direct integration with up to 26 devices (Gupta et al., 2021).

6. Distributed computing, cluster systems, and educational platforms

The distributed-systems papers present Raspberry Pi as a physically small but conceptually rich substrate for clusters, fog nodes, and educational consensus networks. A simple eight-node Raspberry Pi 3 cluster yields 32 cores in total, but the paper’s key result is not raw throughput; it is the hard limit imposed by 100 Mbps Ethernet. For a 3000×3000 double-precision matrix multiplication, the estimated time to transfer the matrix and result over the network is

$_2$6

which completely dominates local compute times of 0.196 s with one thread and 0.062 s with four threads. By contrast, Monte Carlo $_2$7 estimation,

$_2$8

scales well because communication is negligible; the paper reports linear or near-linear speedup when using up to three cores per Pi, but degraded efficiency when all four cores are saturated, likely because OS-level processes contend for resources (Cicirello, 2017).

A more service-oriented variant appears in "HPC as a Service: A naive model" (Imran et al., 2020), where Raspberry Pi 3B and 3B+ nodes run Raspbian Stretch, MPICH, and MPI4py, while another Raspberry Pi 3B+ hosts Apache, PHP, MySQL, and Bootstrap to expose parallel jobs through a browser. Users upload Python plus MPI4py code, specify process count, and obtain output via the web interface. The authors’ Monte Carlo $_2$9 benchmark shows that execution time decreases significantly as process count increases, illustrating that an RPi-based Beowulf cluster can be wrapped as a cloud-like service without exposing SSH, MPI configuration, or node management to end users (Imran et al., 2020).

Con-Pi extends the same hardware class into distributed edge and fog computing. Built on Raspberry Pi 3 Model B+, it decomposes each node into PiController and PiService, uses Flask REST services and Docker containers, and makes energy-aware offloading decisions based on battery state, solar harvesting, and peer discovery. The total per-node energy is modeled as

$^{222}$0

and the objective is to maximize $^{222}$1, the operative lifetime of the most constrained RPi. In real-world experiments, Con-Pi improves response time, RAM usage, and battery management relative to FogBus and FogPlan, and a pest bird deterrent system demonstrates container migration among PiJuice- and solar-powered RPis while sensing and actuation remain local (Mahmud et al., 2021).

Educational and tangible-computing systems occupy the same space. A private Ethereum Proof-of-Work network running on Raspberry Pi 4 or 5 boards turns each Pi into a full node plus miner with its own LCD, while a master Pi provides a web-based interface to configure difficulty, peer topology, latency, and CPU core allocation, making forks, uncle blocks, and consensus degradation directly visible (Ivashkevich et al., 25 Nov 2025). In "Building music with Lego bricks and Raspberry Pi" (Barbancho et al., 2024), three separate RPis implement a synchronism box, beatbox, and melody box; the melody subsystem uses resistor-encoded Lego towers with

$^{222}$2

a voltage divider,

$^{222}$3

and MCP3008 ADCs to convert brick height into MIDI notes, while the whole system is synchronized by MIDI clock up to 208 BPM (Barbancho et al., 2024).

7. Limits, misconceptions, and directions opened by the literature

Taken together, the cited studies suggest that the common reduction of Raspberry Pi to a hobbyist-only platform is inaccurate. Within this corpus alone, RPis act as standards-compliant ETSI C-ITS stations, Lost-In-Space star sensors, ImageNet-scale training nodes for incremental local SGD on deep features, and network-facing slow-control gateways in a 2 PFLOPS HPC cluster (Farina et al., 19 Mar 2025, P et al., 2022, Do, 2022, Otte et al., 2018). At the same time, the literature is equally clear that these uses succeed only when the system design respects the board’s bottlenecks.

One repeated misconception is that nominal model compactness or FLOP count alone predicts edge performance. The papers directly contradict that view. TripleNet-S remains faster on Raspberry Pi 4 than several competing lightweight CNNs despite having about 4G FLOPs, because its connection pattern and memory behavior are more favorable on the Pi; conversely, unstructured magnitude pruning under TensorFlow Lite does not reduce model size or latency because the runtime still uses dense kernels (Ju et al., 2022, Ameen et al., 2023). The operational implication is that cache behavior, concatenation overhead, delegate support, and runtime format can matter more than FLOPs or sparsity percentages.

A second misconception is that Raspberry Pi GPIO should directly absorb every field I/O duty. The HPC monitoring paper reports unstable behavior when four DS18B20 sensors were directly connected to a Raspberry Pi 2B, requiring power-cycling, and therefore moved sensor I/O behind an Arduino for reliability; PiEEG requires that the Raspberry Pi be powered from a 5 V battery rather than the mains for safety during human-body measurements; the C-ITS station reaches about 80 °C under sustained CAM transmission and benefits from added cooling; the eight-node Pi cluster is fundamentally constrained by 100 Mbps Ethernet; and StarberrySense, although accurate and power-efficient, is not radiation-hardened and therefore aimed at short-duration LEO missions rather than long-life space avionics (Otte et al., 2018, Rakhmatulin et al., 2022, Farina et al., 19 Mar 2025, Cicirello, 2017, P et al., 2022).

Future work proposed in the papers is correspondingly pragmatic. The PiEEG work suggests moving more of the online algorithm to C or C++ to reduce latency and enabling more advanced BCI paradigms such as P300 and ERPs (Rakhmatulin et al., 2022). The Raspberry-Pi deep-learning optimization study points toward structured pruning, quantization-aware training, and runtime/compiler support that can actually exploit sparsity on ARM (Ameen et al., 2023). The ImageNet training paper proposes a distributed incremental local SGD implementation on a Raspberry Pi cluster (Do, 2022). TripleNet invites more hardware-aware architecture design for ARM memory systems (Ju et al., 2022). StarberrySense explicitly proposes an upgrade from Raspberry Pi Zero to Raspberry Pi Zero 2 and improved optics to raise update rate (P et al., 2022). The aggregate picture is not of a universal replacement for workstation-, server-, or rad-hard hardware, but of a research platform whose unusually favorable ratio of cost, software openness, and peripheral access repeatedly makes new classes of experiment feasible.