AI-IoT Smart Wheelchair System

• AI-IoT based smart wheelchair systems are integrated platforms combining assistive robotics, multi-modal sensing, and IoT connectivity to enhance user mobility.
• They employ modular architectures with edge computing, sensor fusion, and cloud telemetry to support autonomous navigation, health monitoring, and multi-modal control.
• Research advances focus on robust visual SLAM, BCI-based intent detection, and secure remote updates to improve assistive mobility solutions.

An AI-IoT based smart wheelchair system is an integrated assistive robotics platform designed to augment or replace conventional mobility aids by fusing artificial intelligence, multi-modal sensing, and Internet of Things (IoT) connectivity. These systems target users with mobility or communication impairments, providing context-aware autonomous mobility, robust health monitoring, multi-modal user interfaces, and remote supervisory capabilities. Research in this domain combines advances in computer vision, control systems, sensor fusion, human–machine interaction, and edge–cloud architectures, as evidenced by modular prototypes and literature covering shared-control architectures, reinforcement learning, brain–computer interfaces (BCIs), deep learning–based perception, and standards compliance (Ramaraj et al., 2022, Kim et al., 2023, Sarkar et al., 6 Jan 2025, Leaman et al., 2017, Islam et al., 17 Jan 2026, Hossain et al., 6 Jan 2026).

1. System Architecture and Hardware Integration

Modern AI-IoT smart wheelchair systems are architected with a modular philosophy, distributing computational and sensor responsibilities between edge devices, embedded controllers, and cloud resources. A representative block diagram partitions functionality across several layers: perception (sensor fusion, SLAM), decision-making (planning, user intent), control (actuators, safety), and IoT middleware (telemetry, OTA updates, remote monitoring). Data flows hierarchically: raw sensor data are processed at the edge, decision modules generate actuation commands, control laws close low-level loops, and the IoT layer ensures bidirectional data exchange with cloud dashboards and caregiver terminals (Leaman et al., 2017).

Common hardware modules include:

Module	Typical Components	Functionality
Perception	Stereo/RGB-D camera, LiDAR, IMU, EEG headset, biosensors	Visual odometry, SLAM, BCI signal acquisition
Control	Differential-drive DC/BLDC motors, H-bridge, encoders	Locomotion, velocity/torque regulation
User Interface	Joystick, gesture glove, speech mic, eye-blink detection	Multi-modal user control
Health Monitoring	ECG, SpO$_2$, temperature, accelerometers (fall detection)	User safety, vital sign tracking
IoT/Edge Compute	ESP32, Jetson Nano, Raspberry Pi, cloud links (MQTT, HTTP)	Edge inference, telemetry, OTA update

Systems such as those proposed by (Ramaraj et al., 2022) and (Islam et al., 17 Jan 2026) utilize modular mounting of compute units (Jetson Nano or ESP32), a standard DC motor drive base, and camera/LiDAR suites for perception. Autonomy-enabling sensors replace or complement traditional wheel odometry; for example, (Ramaraj et al., 2022) achieves visual odometry solely via stereo camera and IMU, omitting wheel encoders and reducing retrofitting complexity.

2. AI Algorithms for Navigation, Control, and Intent Detection

AI-IoT wheelchairs deploy a diverse array of algorithms for localization, mapping, navigation, user intention, and environmental awareness.

• Visual SLAM and Odometry: The RTAB-Map package under ROS constructs a pose graph $G = (V, E)$, where nodes encode camera poses and edges encode sequential or loop-closure constraints. Optimization aims to minimize weighted pose error: $$\min_{x} \sum_{(i,j)\in E} \|\Omega_{ij}^{1/2} (Tr_{ij} \ominus (x_i^{-1} x_j))\|^2$$ (Ramaraj et al., 2022).
• Shared-Control and Arbitration: Command blending achieves seamless transitions among manual, semi-autonomous, and full-autonomous modes. A typical blending law is $u = (1-\alpha)u_\text{user} + \alpha u_\text{auto}$, with $\alpha$ determined as a function of user deviation from the planned path (Ramaraj et al., 2022, Leaman et al., 2017). Reinforcement learning (Q-learning) and POMDP frameworks further adapt blending and prompting dynamics (Kim et al., 2023, Leaman et al., 2017).
• Intention and Command Detection:
  • BCI: Mu rhythm features are classified using sparse Bayesian extreme learning machines (ELM) for intent extraction, while event-related potentials (P300/SSVEP) are detected via band-specific spectral and template-matching methods (Kim et al., 2023).
  • EMG/Gesture: LSTM models or threshold-based classifiers decode muscle or accelerometer signals for gesture commands (Kim et al., 2023, Hossain et al., 6 Jan 2026, Islam et al., 17 Jan 2026).
  • Hybrid interfaces: Electrooculography (EOG), speech, and joystick signals are arbitrated per a priority ladder and fused for robust input (Hossain et al., 6 Jan 2026).
• Path Planning and Obstacle Avoidance: Systems rely on classical (A* global, dynamic window approach local) and learning-based planners, costmap inflation for collision margin, and deep learning–based object detectors (e.g., YOLOv8) for semantic perception (Ramaraj et al., 2022, Islam et al., 17 Jan 2026). In dual-layer architectures, ultrasonic sensors handle immediate proximity hazards while vision-based networks detect distant or occluded obstacles.
• Health and Anomaly Detection: Continuous monitoring with biosensors is processed via digital filtering and calibrated against clinical standards. Custom anomaly detection, often cloud-triggered, supports alerting for vital sign deviations, falls, and ambient hazards (Hossain et al., 6 Jan 2026, Islam et al., 17 Jan 2026).

3. Multi-Modal User Interfaces and Human–Machine Interaction

AI-IoT smart wheelchair systems leverage redundant and complementary user input modalities to maximize accessibility:

• Joysticks and Touch: Classical control with dead-zone and linear mapping to velocity/turn. Achieved up to 99% command accuracy in 500-trial evaluations (Hossain et al., 6 Jan 2026).
• Gesture Control: Wearable gloves or accelerometer modules, threshold-classified or processed via machine learning; 95.5%–95% recognition in controlled trials (Islam et al., 17 Jan 2026, Hossain et al., 6 Jan 2026).
• Eye-Blink/EEG: Single and double-blink events (e.g., blink strength $B[n] \geq B_{th}$, double blink within $\Delta t \leq 0.4$ s) trigger direction commands. Primary input for users with motor impairment (Sarkar et al., 6 Jan 2025), integrated via Android/Arduino/Bluetooth in clinical prototypes.
• Speech: Smartphone-based speech recognition with BLE transmission, 97 ± 2% accuracy, 20 ms latency from utterance to motion (Hossain et al., 6 Jan 2026).
• BCI: Intent detection via mu, P300, SSVEP, with state-of-the-art ELM, RQNN filtering, and adaptive RL for validation (Kim et al., 2023).
• Arbitration: Hierarchical and adaptive arbitration logic ensures fail-safe operation by prioritizing safety interrupts, hardware buttons, and override controls (Hossain et al., 6 Jan 2026).

Continuous efforts are documented to improve inclusivity (e.g., support for emotional state integration (Kim et al., 2023)), privacy-preserving data handling, and reduction of user training overhead via personalized, curriculum-based RL.

4. IoT Connectivity, Edge–Cloud Integration, and Telemetry

Smart wheelchairs employ IoT networks for real-time supervision, fleet coordination, and health analytics. Architectures routinely feature:

• Edge Compute and Local Arbitration: Devices such as Jetson Nano or ESP32 process sensory and command data, with secondary MCU for time-critical motor and safety control (Ramaraj et al., 2022, Hossain et al., 6 Jan 2026).
• Networking: Wi-Fi (802.11), BLE, Zigbee, LoRaWAN for device interlinking, smartphone pairing, and connection to smart-home hubs. Security protocols include AES-128/CCM, TLS (Kim et al., 2023, Hossain et al., 6 Jan 2026).
• Cloud Services: Data and telemetry are uploaded (e.g., via MQTT to AWS IoT Core, HTTP POST to ThingSpeak), with periodic cloud-based retraining of AI models (ELM weights, RL policies), remote OTA parameter updates, and fleet-wide map/database sharing (Ramaraj et al., 2022, Kim et al., 2023, Hossain et al., 6 Jan 2026).
• Caregiver and Monitoring Dashboards: Real-time Android/web dashboards display vital signs, offer remote configuration, and trigger notifications (SMTP, SMS API) on events such as falls or abnormal biosignals (Hossain et al., 6 Jan 2026, Islam et al., 17 Jan 2026).
• Latency and Energy Budgets: Closed-loop control latencies of 20–22 ms and idle power draws of ~87 mW are attainable using contemporary MCUs. Trade-offs between wireless duty cycle and battery life, as well as DSP vs. MCU-based preprocessing, are documented (Hossain et al., 6 Jan 2026).

5. Experimental Performance, Validation, and Standards Compliance

Empirical validation spans simulation (Gazebo, AWS Robotics environments) and real-world lab and clinical settings. Reported outcomes include:

Metric	Value	Context / Source
Localization RMSE	0.05–0.08 m	Visual SLAM, Gazebo/lab (Ramaraj et al., 2022)
Path following error (lateral, mean)	~0.07 m	Sim/real (Ramaraj et al., 2022)
Obstacle avoidance (static/dynamic)	100% / 70%	20 trials (Ramaraj et al., 2022)
End-to-end plan latency	SLAM 25 ms, global 100 ms, local 30 ms	(Ramaraj et al., 2022)
Gesture control accuracy	95–99%	400–500 commands (Islam et al., 17 Jan 2026, Hossain et al., 6 Jan 2026)
Speech input accuracy	97 ± 2%, 20 ms	(Hossain et al., 6 Jan 2026)
Health sensor RMSE (HR, SpO2, Temp.)	≤2 bpm, ≤1%, ≤0.5°C	Clinical calibration (Hossain et al., 6 Jan 2026)
Emergency alerts detection	100% (smoke/fall/obst.)	3–5 trials (Sarkar et al., 6 Jan 2025)

System designs prioritize modularity—facilitating isolated upgrades to perception/AI stacks—and scalability, maintaining ROS node-based or microservice-based architectures (Ramaraj et al., 2022, Islam et al., 17 Jan 2026). Standards including ISO 7176-31 and IEC 80601-2-78 are explicitly targeted in certain prototypes (Hossain et al., 6 Jan 2026).

6. Challenges, Solutions, and Research Directions

Technical and implementation challenges include:

• Signal quality and classification: EEG signals are inherently noisy and subject to artifact contamination (EOG/EMG interference). Solutions comprise sparse Bayesian ELM-based feature selection, RQNN filtering, and sensor fusion with IMU, EMG, and context signals (Kim et al., 2023).
• Localization robustness: Visual SLAM may drift, especially under dynamic crowd occlusions. Hybridization with UWB tags, wheel encoder odometry, and periodic loop closure via markers mitigates this (Kim et al., 2023).
• User adaptability and privacy: BCI gestures have a steep learning curve, and users express concern regarding persistent logging. Adaptive RL-based personalization and modular privacy controls (on-device encryption, GDPR-compliant consent) are integration foci (Kim et al., 2023).
• Connectivity disruption: Systems accommodate cloud link loss with fail-safe local autonomy (Leaman et al., 2017).

Proposed and ongoing research avenues include federated meta-RL for personalized navigation, dynamic computation offload (KubeEdge), co-robotic arms with imitation learning for object manipulation, fusion of fNIRS and EEG for higher-throughput intent detection, and integration with smart home automation leveraging full-edge–cloud orchestration (Kim et al., 2023, Ramaraj et al., 2022, Hossain et al., 6 Jan 2026).

7. Outlook and Prospects

AI-IoT smart wheelchair systems represent a modular convergence of autonomous robotics, digital health, and pervasive computing for independent mobility. As documented in the contemporary literature, continued advancements in BCI/adaptive AI, multi-modal interaction, embedded neural networks, and certified IoT frameworks (e.g., MQTT/REST with secure enclaves) will drive enhanced autonomy, safety, and quality of life for users requiring complex assistive technologies. Standardization of control interfaces, robust cloud synchronization, and co-design with end-users and caregivers remain central to future research and deployment (Leaman et al., 2017, Kim et al., 2023, Islam et al., 17 Jan 2026, Ramaraj et al., 2022).