Page 9 - Read Online

Page 334 Zhang et al. Intell. Robot. 2025, 5(2), 333-54 I http://dx.doi.org/10.20517/ir.2025.17

1. INTRODUCTION
Strokeis recognizedasthethirdleadingcauseofdisability anddeath worldwide, accordingtotheWorldStroke
[1]
[2]
Data Showcase . The current global incidence of stroke has surpassed 15 million cases . Furthermore, the
number of new stroke cases each year is extremely large and shows a continuously increasing trend. The ma-
jority of stroke patients experience severe disabilities, predominantly motor movement disorders of the limbs,
[3]
which significantly affect their daily living activities . To promote functional recovery and prevent further
deterioration, rehabilitation is considered crucial for these patients. The recovery of motor function in stroke
patients mainly depends on early intervention and continuous rehabilitation training, including physical ther-
apy (PT), occupational therapy (OT), and robot-assisted training. These methods aim to promote neural
plasticity and functional reconstruction. Among them, repetitive training has been proven to be an effective
[4]
means of enhancing muscle strength and improving motor control . Compared with traditional PT and OT,
robot-assisted movement training has shown significant advantages in clinical outcomes and biomechanical
measurements, and its effectiveness in promoting the rehabilitation of stroke patients has been widely con-
[5]
firmed . Robot-assisted training provides consistent and precise repetitive movements, allowing patients to
perform high-intensity repetitive training, which is difficult to achieve in traditional PT and OT. This training
method can provide mechanical assistance to the limbs, enabling patients to perform repetitive training of
voluntary movements, which is of great help to patients who have difficulty moving automatically due to the
aftermath of a stroke. As a nascent technology, robot-assisted training has demonstrated its distinctive merits
in clinical and biomechanical assessments, providing new avenues for the rehabilitation of stroke patients [6,7] .

Due to the intimate physical interaction between exoskeletons and humans, motion planning is an essential
component of exoskeleton control systems. Trajectory planning is not only related to the wearer’s safety and
comfort but also directly affects the effectiveness of rehabilitation training. Trajectory planning algorithms
can be primarily divided into two major categories: Cartesian space trajectory planning and joint space trajec-
tory planning. Cartesian space trajectory planning focuses on the motion trajectory of the end-effector in the
Cartesian coordinate system, while joint space trajectory planning concentrates on the motion trajectory of
the robot’s individual joints in the joint space. Trajectories in the task space need to be converted from Carte-
sian coordinates to joint angles using inverse kinematics [8,9] , but it is difficult to obtain a non-unique solution
for inverse kinematics in the human-machine workspace [10,11] . Optimization-based trajectory planning meth-
ods can systematically consider various performance indicators, such as time, energy consumption, and path
length, in search of the theoretically optimal solution. Common optimization models include the minimum
jerk model [12,13] , minimum torque model [14,15] , minimum inertia model [16,17] , minimum potential energy
model [18,19] , etc. Wang etal. utilized anadaptivefrequency oscillator(AFO) toextract high-level featuresfrom
the active arm movement, then combined the motion rhythm with the principle of minimum jerk to generate
an optimal reference trajectory, which is synchronized with the patient’s movement intention and the move-
ment patterns of healthy individuals [20] . Sampling-based methods [21,22] , such as rapidly-exploring random
trees (RRT) and probabilistic roadmaps (PRM), are suitable for complex environments and high-dimensional
spaces, capable of handling narrow or cluttered obstacle situations, but may require a large number of sam-
ples to ensure the feasibility of the path, leading to high computational costs. Machine learning-based tech-
niques [23–26] and learning methods based on Hidden Markov Models (HMM) and Gaussian Mixture Models
(GMM) [27] can learn from data and adapt to different tasks and user behaviors, but they require a substantial
amount of training data, and their generalization capabilities for unseen situations may be limited.

The rehabilitation robot is designed to assist the patient in moving along a specific trajectory. The objective of
the rehabilitation robot is to minimize tracking errors, thereby facilitating the patient’s acquisition of normal
movement patterns. This training modality is particularly well-suited for patients exhibiting impaired muscle
movement during the pre-rehabilitation phase, as it serves to avert muscle atrophy. However, the implementa-
tion of passive rehabilitation training requires caution to ensure the safety of the patient and the effectiveness
of the training. Therefore, the establishment of motion-restricted areas is particularly important in passive