Table 3.

Brain-computer interface technologies for adaptive assistive mobility devices.

Technology name (type): additional sensors	Machine learning tools	Contributions	Drawbacks
TDS-iPhone-PWCa (haptic): magnetic sensors [46]	Sensor signal processing algorithm	An alternative USIb for people with spinal cord injury or upper limb paralysis	Tongue piercing can be a painful and uncomfortable option for some users. Extensive training is required for calibration.
Intelligent smart walker (haptic): force or torque sensor [49]	N/Ac	An intuitive rule-based speed controller for a smart walker	Young and healthy subjects were used, so the result is not a true representation of the typical users of the walker.
EyeCane (CVId, haptic, and auditory): infrared emitters, auditory frequency actuator, and tactile actuator [52]	N/A	Low cost, lightweight, small and easy to use electronic travel aid for distance estimation and navigational assistance, long battery life (one whole day), intuitive to the user, and short training time (<5 minutes)	Only an indoor experiment was conducted.
Electronic mobility cane (CVI, haptic, and auditory): liquid detection, 6 ultrasonic sensors, a metal detector, a microvibration motor, and a mono earphone [51]	A novel algorithm named way-finding with reduced information overload.	Offers real time multiple obstacle detection and way-finding assistance simultaneously to patients with visual impairments by an auditory (voice message) and tactile (vibration) feedback	Extensive training time (20 hours); the cognitive and perceptual load has not been ascertained
Jet Propulsion Laboratory BioSleeve (haptic): electromyography and IMUe sensors [44,45]	A multiclass support vector machine classifier	Intuitive control of robotic platforms by decoding as many as 20 discrete hand and finger gestures	Has not yet been integrated and tested with assistive mobility aids to determine its applicability
Smart cane (haptic): IMU and FSRf sensors [48]	C4.5 decision tree, artificial neural network, support vector machine, and naive bayes	To monitor and distinguish between different walk-related activities during gait rehabilitation	Fall and near-fall detection was not considered in its design and implementation.
An ARTAg power wheelchair platform (CVI and haptic): haptic controller, laser scanner, SICK laser measurement, and IMU sensor. [50]	Gaussian process regression model	Implementation of a learned shared control policy from human-to-human interaction	The efficiency of the learning process is dependent on the human assistant, who is prone to errors and might miss out on the certain intent of the user.
Multiple controlled interfaces smart wheelchair (haptic and auditory): microphone, joystick, leap motion, and ultrasonic sensor [53]	An algorithm for the control and execution of commands	Multiple control interfaces	Lack of details on the performance of each interface and limited testing scenarios
MyoSuit (haptic): IMU sensor and two electric motors [47]	N/A	Lightweight, soft wearable robot to aid users with a level of residual mobility during locomotion tasks	Only one incomplete spinal cord injury participant was selected for testing, so it is difficult to validate its performance.

^aTDS-iPhone-PWC: tongue drive system to iPhone electric-powered wheelchair

^bUSI: user system interface.

^cN/A: not applicable.

^dCVI: computer vision interface.

^eIMU: inertial measurement unit.

^fFSR: force sensitive resistor.

^gARTA: assistive robotic transport for adults.