IEEE Robotics & Automation Magazine - March 2023 - 30

robust: the user has full authority over the wheelchair when
the input is safe and benefits from progressive assistance during
difficult maneuvers (e.g., reversing out of an elevator).
MODEL-BASED SHARED CONTROL
While the dynamics of a power wheelchair can be precisely
characterized and have been widely used for control design in
the literature, it is challenging to combine the capabilities of a
machine (as described by its dynamic model) with the limited
unpredictable information coming from 1) the environment
and 2) the human user (e.g., the joystick interface is a projection
of the user's intention). To compensate for the limited and
incomplete information available from real-time (online) measurements,
recent research has explored stochastic models.
One way of implementing a stochastic model is to use probabilistic
shared control [25]. However, this technique incurs
considerable computational cost to generate possible wheelchair
trajectories, which may preclude its use in real-time
applications. To circumvent this limitation, in [26], UCL's
group proposed to use stochastic dynamic programming
(SDP), which takes all the computation burden offline. The
outcome is a lookup table that can be readily used online by
the assist-as-needed algorithm. More specifically, in [26], a
model-based control architecture has been introduced to solve
the obstacle avoidance problem. It consists of four blocks,
where some are deterministic and others are stochastic.
First, the wheelchair dynamics block includes the physical
equations of motion of a two-wheeled differential drive vehicle
(for the experimental model identification, see [16]). Second,
the environment block is used to model the static obstacles,
with the vehicle having limited knowledge of the global map.
In fact, a local map around the wheelchair is built based on the
distance measurements coming from an array of sensors (e.g.,
ultrasonic or laser sensors). Third, the driver intention block
includes stochastic models of driver intention (e.g., an expert
driver capable of maneuvering the wheelchair at high speed
yet seldomly hitting obstacles, a " blind " driver for which the
probability of hitting obstacles or avoiding them is the same,
and a naughty child who intentionally advances at high speed
toward the obstacles with the intention to hit them as a learning
experience). Fourth, the supervisory control block computes
optimal assist-as-needed actions specifically tailored to
each driver model, which comes in the form of multidimensional
lookup tables.
HUMAN-MACHINE INTERACTION
AND HAPTIC FEEDBACK
The ADAPT project gave the French and English teams a
great opportunity to design innovative human-machine interfaces.
Among other devices, haptic interfaces have been conceived
to assist users with wheelchair's navigation.
Two types of systems have been tested: a haptic joystick
[24] and a vibrotactile armband [27] (see Figure 8). Both
devices can be easily interfaced with the control system of any
consumer-grade wheelchair. While driving the wheelchair, a
reactive force is applied by the haptic joystick to the hand of
the user. By offering resistance in the direction of an obstacle,
the user is thus indirectly informed about the safe trajectory to
follow. However, the haptic joystick remains a simple decision
support system which does not replace the driver, who is in
full control of the wheelchair at all times. The armband can
be worn anywhere on the upper and lower limbs, depending
on the user's sensory capabilities. The armband is composed
of four evenly spaced vibrotactile actuators, powered by a
lithium-ion battery and controlled by an embedded wireless
electronic board. The armband is inexpensive and provides
intuitive commands (information about the path to follow
or about the presence of obstacles, in the form of a direction
with respect to the current orientation of the wheelchair). As a
result, users do not need long training sessions.
The sensor-based shared control algorithm developed at
Vibrotactile
Armband
Wheelchair
Joystick
Controller
Ultrasonic
Sensors
INSA Rennes is compatible with the haptic feedback provided
by the haptic joystick and the vibrotactile armband. The feedback
is computed by processing range measurements coming
from the wheelchair, and it supports the user during spatial
navigation tasks. The haptic feedback can be employed in conjunction
with the algorithm in [24] (progressive assistance while
approaching an obstacle) or stand-alone (i.e., the control is not
delegated, and the user has full authority over the wheelchair).
FIELD TESTS AND CLINICAL TRIALS
FIGURE 8. A volunteer wears a vibrotactile armband developed at
INSA Rennes on his right upper arm while driving UCL's wheelchair.
30 IEEE ROBOTICS & AUTOMATION MAGAZINE MARCH 2023
CLINICAL EVALUATION
Experiments and regular roundtable sessions with patients,
robotics experts, occupational therapists, and specialists in
rehabilitation medicine have played a key role throughout the
ADAPT project. In fact, if the former were necessary to validate
the robotic solutions developed, the latter were essential
to ensure that the specific needs of the patients were satisfactorily
met. The research ethics committee approved the
IEEE Robotics & Automation Magazine - March 2023

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - March 2023
