Dynamic Shared Control For Human Wheelchair Cooperation Engineering Essay

Published: Last Edited:

This essay has been submitted by a student. This is not an example of the work written by our professional essay writers.

Abstract-Shared control is a common used method for human and wheelchair cooperation. However, most of the previous shared control methods didn't think much of the effect caused by the difference in the user's control ability. The control weight of a user in these methods is irrelevant to the user's capability or the driving conditions. In this paper, a dynamic shared control method is proposed to adapt wheelchair's assistance to the variations of user performance and the environmental changes. Three evaluation indices including safety, comfort and obedience are designed to evaluate wheelchair performance in real time. A minimax multi-objective optimization algorithm is adopted to calculate the user's control weight. The results of lab experiments and elderly home field tests show that this method can adapt the degree of wheelchair's autonomy to the user's control ability and it makes driving wheelchair much easier for elder people.


As the proportion of elderly population has increased consistently over the past decades, smart wheelchair based on the traditional mobile robots technology is developed to help old and disabled people with their daily lives. By using a smart wheelchair, people could get rid of onerous maneuver and dangers like collision and fall. The enhancement of independent mobility by a smart wheelchair can also help in rebuilding people's confidence of social skills. Since people's control ability could be affected by various factors, such as inherent control ability, fatigue, environmental visibility and degree of crowding, it varies according to human's physical state and the environmental condition. In order to respect user's self-esteem and make less conflict with user, the control system of a smart wheelchair should change the degree of autonomy according to the user's control ability. This user adapted idea is very useful especially when people are having rehabilitation training. In this paper we proposed a method that can evaluate the user's control ability from the user's commands and adjust the user's control weight according to his or her control ability.

Numerous shared control schemes had been proposed since 1990s [1]. These shared control schemes can be divided into two categories according to the control level a user takes part in: behavior level sharing and planning level sharing.

Behavior level sharing is a commonly used method during the early stage of the smart wheelchair. There are usually two ways for a wheelchair to cooperate with a human. The first way is that a wheelchair goes towards a direction that a user pointed out by joystick and the assistive system provides some obstacle avoidance algorithm to ensure safety [7], [12]. In the second way, the user's commands are treated as a behavior which is executed with other autonomous behaviors (e.g. obstacle avoidance behavior, wall follow behavior). One research topic in this level is how to choose the proper behavior when the wheelchair is facing different conditions. These behaviors could be executed in parallel [3], or selected via probabilistic methods [8]. In this level the user's control ability is assumed to be the same and the control weight of a user is constant. Implements of behavior level sharing are intuitive and fast, but drawbacks like oscillation, local minimum and large control magnitude change are prevalent.

Planning level sharing takes the user's intention into account while doing planning. The wheelchair follows orders coming from a planner, and user expresses his or her control intention by moving a joystick. When the user's control intention conflicts with the planner's order, the control system will modify the user's command [9] or re-plan the task [5], [10]. The user's intention of doing a certain task (e.g. door passage) is measured by defining intention prediction functions [5]. When a user is doing a pre-defined task, the value of one of the prediction functions will rise above a threshold and the planner will order the wheelchair to follow a pre-defined trajectory correspond with the selected function. Statistical methods are also used in this level. In one of these methods a sub-goal the user wants to reach is estimated by a Bayesian Network and the user's commands. When the estimated sub-goal differs from the pre-planned sub-goal, the planner will replan the trajectory [10]. Although the user's intention is concerned in this level, he is forced to obey the trajectory given by the planner. In other words, the variation of people's control ability is not carefully considered by this kind of shared control.

A new kind of shared control method had been proposed by Poncela[2] and Urdiales [11]. This method is developed from the traditional reactive control idea. It defines an efficiency function to evaluate the user's control ability and adjusts the user's control weight according to the function value.

The system designed in [2] has taken the user as an integral part of the wheelchair system and the idea of evaluation function is notable. Instead of using the evaluation function directly we take the weight adjustment problem as an optimal problem. By finely defining some indices for wheelchair performance, the control weights of a user and a machine can be figured out by solving a multi-objective optimization problem.

Fig. 1 Hardware structure of the JiaoLong wheelchair.


The JiaoLong wheelchair prototype is based on an ordinary electric wheelchair. The wheelchair is equipped with mobile robot sensors including LMS200 laser range finder (LRF), camera, odometry etc. An on-board computer (OBC) running Windows XP is used to implement the proposed shared control algorithm. A Smart Motion Controller (SMC) based on a DSP processor is adopted to execute motion control commands for the wheelchair.

As shown in Fig. 1, the system consists of three units: 1) Sensor unit, 2) Control unit, 3) Wheelchair unit. The sensor unit, including two encoders, a LRF and a camera, is used to detect the environmental information near the wheelchair. The control unit, including OBC and SMC, is used to process sensor information and execute control algorithm. The wheelchair unit, including motors, battery and joystick, is a combination of the wheelchair's basic components. A user gives a desired direction and speed to SMC via joystick, SMC then passes these orders to OBC via a serial port. The OBC running the main control procedures calculates motion command according to the user's inputs and the sensor information, and it also provides a human-friendly interface. This entire system is powered by two 12 V lead-acid batteries connected in series. The maximum translational and rotational speeds of the wheelchair are 500 mm/s and 50 deg/s. Considering the inconvenience of these users with dyskinesia, this system also provides human-machine interfaces via voice control, gestures control, EEG control, etc.

Shared Control With Weight Optimization

System Architecture

Humans, especially old or disabled ones, are less precise in maneuver, do not preserve curvature well and sometimes have difficulty in perceiving their surroundings. Pure reactive controller has problems like fall in local traps, oscillation and unsmooth trajectory. However, human is always good at high level planning, and machine is precise in detecting environmental information and executing motion control. It is significant to find out a method that can combine human control ability and machine control ability effectively. Machine assistance should be adaptable to the difference of the users' control abilities. A Dynamic Shared Control (DSC) algorithm rooted in this idea is proposed in this paper.

Fig. 2 depicts the functional diagram of this algorithm. vuser and ωuser are the desired translational and rotational speed

Fig. 2 Architecture of the dynamic shared control system.

from user. ωmach is the desired rotational speed generated by the reactive controller. ωoptimal is the desired rotational speed generated by the weight optimizer. vfinal and ωfinal are the desired translational and rotational speed to be sent to SMC. Obs_Info is the obstacle distribution map obtained from the LRF data and the odometry data. v0 and ω0 recorded by the odometry are the current translational and rotational speed of the wheelchair.

There are two key parts in this architecture: the reactive controller and the weight optimizer. The reactive controller provides basic machine help using MVFH&VFF methods [3]. The weight optimizer assigns control weights to user and the reactive controller. The minimum vector field histogram (MVFH) method and vector force field (VFF) method which the reactive controller adopts are developed by the University of Michigan [3], [4]. Instead using sonar the MVFH&VFF here use a LRF to perceive obstacles. Obs_Info records the positions of obstacles around the wheelchair. This obstacle information are detected by the LRF and recorded along with the odometry data. Details of the weight optimal algorithm will be discussed in the following section. vuser and ωuser are generated by joystick displacement.

This algorithm includes five steps: 1) Update an obstacle distribution map (Obs_Info) from the latest LRF data and odometry data. 2) The reactive controller calculates the machine output ωmach by using MVFH&VFF. 3) The weight optimizer calculates the rotational speed ωoptimal by using the proposed algorithm which will be discussed later. 4) The motion controller calculates vfinal and ωfinal according to vuser and ωoptimal. When the distance between the wheelchair and the nearest obstacle reduces to a certain value (0.5m in current experiments), the motion controller will reduce vfinal according to the obstacle distance. When the obstacle distance is above the threshold, vfinal will always be equal to vuser. 5) Send vfinal and ωfinal to SMC via RS232.

The control cycle of this algorithm is 200 ms which is mainly restricted by the sampling period of the LRF.

Weight optimization

As mentioned above, the user's performance in maneuver is measured by some evaluation indices and the weight optimizer calculates the user's control weight according to his performance. So the definition of evaluation indices is the base of this algorithm. According to the results of the questionnaire surveys at an elderly home, we propose three indices: safety, comfort and obedience to evaluate the wheelchair's performance. safety is used to measure the probability of collision. comfort is used to measure the smoothness of velocity. obedience is used to measure the degree of obedience to the user's control intention.

In order to facilitate the calculation of the user's control weight, we tend to define the indices as functions of wheelchair movement and obstacle information, i.e. index = index (Obs_Info, Chair_Mvt). Once these evaluation indices are defined, the shared control problem can be written in a standard multi-objective optimization problem form (Eq. (1)).


Where, κ1 and κ2 represent the control weights of the user and the reactive controller; v(t) and ω(t) represent the translational and rotational speed to be sent to the SMC. To avoid violent change in motion command, we define ω as a linear combination of ωuser and ωmach. This equation means finding the user's control weight is equivalent to finding the proper κ1 and κ2 to maximize the three max(∙) items under the restrictions stated after s.t.. As the translational velocity in MVFH is equal to vuser(t) as long as there is no possible collision, we restrict v(t) to be equal to vuser(t).

Evaluation Indices

The safety index needs to be able to reflect the possibility of a wheelchair colliding with obstacles. safety in this paper is defined as:


where, α is a constant. dis measured in millimeter represents the distance between the wheelchair and the nearest obstacle in its path. Fig. 3 is the illustration of dis. The predicted path of the wheelchair is calculated according to the wheelchair's kinematic model. dis is determined by the current and desired velocity of the wheelchair and the obstacles around the wheelchair. Since kinematic model is inaccurate in predicting long-term movement, we only predict the path for the next 4 seconds. The negative exponential function is used to normalize indices, and make the comparison between those three indices convenient.

Experiments show that frequent change of velocity will make user feel uncomfortable. High jerkiness of a wheelchair may also cause danger for people with spinal cord injuries. Therefore, we decide to use a function of velocity change to

Fig. 3 Distance to the obstacles in the predicted path.

evaluate comfort index. Like the definition of safety, we adopt the negative exponential function to normalize comfort.


Where, β is a constant. Since the wheelchair's translational speed is given by user directly and it won't be changed unless a collision is about to happen, there is no component reflects translational speed change in Eq. (3).

The obedience index is used to evaluate the proximity between the user's control intention and the final motion command. As an assistant device it is essential to provide assistance that is congenial to the user's control intention. For the sake of computational simplicity, we use the user's target orientation issued via a joystick to represent the user's control intention. We assume that the closer the final desired orientation and the user's target orientation is; the more obedient the final motion command is. Hence the obedience index is calculated as


Where, ξ* is the orientation calculated from the user's input vuser and ωuser ; ξ is the orientation determined by v and ω; γ is a constant. This index can make the wheelchair moving under the user's intention as long as he is able to maintain safety and comfort.

Simplified optimization algorithm

These three indices are usually contradictory to each other under normal circumstances. For example when a wheelchair is traveling through a crowd, it will be required to turn a big round for safety, but for comfort it is not allowed to do that. Therefore, there is no absolute optimum solution for Eq. (1). An evaluation function of this problem is needed to achieve an effective solution.

It was found that increasing a certain index which is already above a certain value will make the other two indices drop drastically. For example, enforcing safety to increase when it is already above 0.9 will make the wheelchair always choose the most spacious path, and the user will feel the wheelchair is not moving under his control at all. Therefore, a principle we proposed of solving this problem is: always improve the smallest index among the three. In accordance with this principle we choose the minimax method to simplify Eq. (1) to a single objective problem (Eq. (5)).


The algorithm's adaptation could be improved by using this minimax method. The precedence relation among indices will change naturally when facing different situations. For example, when a user is cruising in a spacious room with a wheelchair, the possibility of hitting an obstacle is small, i.e. safety has a high value. In this case max(min(safety, comfort, obedience)) = max(min(comfort, obedience)), then comfort and obedience become priorities. Whereas when a user is passing through the crowd, the possibility of a collision may significantly increase. Then safety could be taken into consideration as a priority since max(min(safety, comfort, obedience)) = max(safety).

A user with good control abilities can drive a wheelchair smoothly and safely, in this situation max(min(safety, comfort, obedience)) = max(obedience), so the wheelchair will be driven completely under the user's will. When the user's control ability drops, he would not be able to preserve smoothness and safety. In this condition, the reactive controller will assist he and its control weight will increase.

Experiment And Discussion

Laboratory Experiments

Both laboratory experiments and elderly home tests were carried out to evaluate this proposed algorithm. In laboratory experiments, volunteers were asked to pretend to have some kinds of disabilities like visual defects, trembling hands. α, β and γ in the three indices are set to 0.0005, 0.006 and 0.003 in the following experiments. These parameters were not very well tuned and determined based on some assumptions. For example, according to the assumption that safety equals 0.63 in case the obstacle distance is 2m, α will be calculated.

Fig. 4 illustrates the experiment in which a user with visual defects tried to pass through a door. The DSC algorithm was enabled in this experiment. The right part of Fig. 4 is the trajectory of the wheelchair. The left part includes four variation curves. From top to bottom, they are user's control weight, user's desired rotational speed, machine's desired rotational speed and the combination of user's and machine's rotational speed. Because of the user's visual defects, he can only perceive the direction to the door but had no idea of the exact position of the door. He could just drive the wheelchair towards the door. At the beginning there was no obstacle around the wheelchair and no collision was possible to happen. The user was capable at that time, obedience became the dominating index, and the user got a relatively high control weight (0.8 in average). At the time point 25, if the wheelchair was still controlled completely under the user's

Fig. 4 The user's control weight in door passage experiment.

Fig. 5 Low control weight of the user with trembling hands

will, there would be a collision at the "possible collision" point (Fig. 4). To preserve safety the DSC algorithm started to decrease the user's control weight. It can be seen from the curve of the user's control weight that it was changed according to his control ability. This seamless cooperation made the movement of the wheelchair safe and smooth.

Fig. 5 illustrates the experiment in which a user with trembling hands tried to move through a corridor. In order to compare with manual operation, the user was asked to drive the wheelchair both with and without DSC. The right part of Fig. 5 is the trajectory of the wheelchair. The left part includes four variation curves. From top to bottom, they are user's desired rotational speed in manual mode, user's desired rotational speed in DSC mode, machine's desired rotational speed in DSC mode and the combination of user's and machine's rotational speed in DSC mode. The user waggled the joystick at a range of about ±25 deg/s. In manual mode the wheelchair was swayed badly. In DSC mode, since the user's maneuver would cause sway and jerkiness, comfort became the dominating index. To make the wheelchair move smooth and comfortable, the DSC algorithm decreased the user's control weight to about 0.2 in average. In this case, the user's low control ability is detected from his jerky joystick movement, then the DSC algorithm decreased his control weight to preserve comfort. The wheelchair's trajectory was

Fig. 6 Evaluation tests at an elderly home.

Fig. 7 Three test tasks.

improved by this decrease.

These above experiments show that the proposed algorithm can adapt the degree of autonomy to the user's control ability. When a user is able to maneuver the wheelchair safely and comfortably, he will get high control weight. When the situation becomes hard for him to drive in, the algorithm will reduce his control weight according to his performance.

Elderly Home Tests

Three kinds of tests were carried out at the Shanghai 3rd Elderly Home (Fig. 6) after numerous laboratory tests.

We invited five elderly people who are in normal states of mind but have some difficulty in mobility. Their ages range from 75 to 84. Two of them have disabilities. One is suffering from weakness in limbs and the other is suffering from left-handed stroke.

All the subjects were asked to drive the wheelchair through three different tasks (Hall Tour, Door Passage, Obstacle Avoidance, Fig. 7) under three operating modes (Manual, MVFH-aided, DSC). It is hoped that we can improve the algorithm by comparing the algorithm performances in different tasks. The wheelchair is equivalent of a common electric wheelchair in Manual mode. Users can get MVFH&VFF avoidance aid in MVFH-aided mode. The DSC algorithm is activated in DSC mode.

Before the tests, each subject was asked to fill out a questionnaire which is used to record his eyesight, mental status, and hands flexibility. After that we will guide him to finish the three tasks in three modes. During tests we record the subject's operational information and the wheelchair trajectory in real time. After finishing each task, the subject was asked to tell us some subjective indicators such as task difficulty, operating difficulty and satisfaction. The average

Fig. 8 The smoothness and fluency of trajectory

Table I Mean and standard deviation of Trajectory Quality



Hall Tour

Door Passage

Obstacle Avoidance










Speed (m/s)

























































training time was about 15 minutes and the test time was about 20 minutes. The wheelchair's speed was about 0.25 m/s in average.

There are two types of evaluation data: 1) Subjective data which records the subjects' feeling of the task difficulty, the operating difficulty and his satisfaction. 2) Objective data which records the task completion time, collision times, trajectory smoothness, trajectory fluency, trajectory length, user's control weight, and the three evaluation indices (safety, comfort and obedience). The definitions of the trajectory smoothness and fluency are shown in Fig. 8. Those nodes in Fig. 8 represent the wheelchair's location at corresponding time points and theτ1 (be set to 80) andτ2 (be set to 1) are proportionality constants. The trajectory smoothness index reflects the curvature of the trajectory and the trajectory fluency index reflects the change of the wheelchair velocity.

The questionnaires show that the door passage task is the hardest one and the hall tour is the easiest. By using the DSC algorithm, user's satisfaction was improved and task like door passage becomes easier.

Table I shows the mean and standard deviation of the speed, fluency and smoothness from tests data of the five subjects. We can see from studying Table I that the average speeds in different tasks have significant differences and the simpler the task the faster the speed is. Speed under DSC mode doesn't win the title due to the reduction of speed when the wheelchair is close to obstacles. It can be seen that the fluency and smoothness indices get a relative high value under DSC mode, meaning that the trajectory quality was improved in DSC mode. These improvements are mainly caused by the restriction of the evaluation indices and the increase of the user's control performance as the tasks difficulty decreased.

To further illustrate the improvement of the wheelchair operation after adopting the DSC algorithm, we plotted trajectories from one subject in three modes (Fig. 9). These trajectories were recorded by odometry which may cause discrepancies between the recorded trajectories and the actual ones. However, these discrepancies could be ignored as long as the trajectory is short. It can be seen in Fig. 9 that the trajectory under DSC mode is smoother than the other two.

Fig. 9 Comparison of trajectories in the obstacle avoidance task

Fig. 10 is a clearance distribution map which shows the distribution of the distances (330 mm~650 mm) between the wheelchair and the obstacles near it under three modes.These distances data were recorded by the LRF and don't contain odometry errors. The minimum clearance in DSC mode is 590 mm, in manual mode the value is 350 mm and in MVFH-aided mode it is 510 mm. This shows that the possibility of colliding with an obstacle under DSC mode is smaller than that under the other two modes.

Conclusion And Future Works

This paper presented a new approach of shared control for semi-autonomous wheelchair. Three indices including safety, comfort and obedience are designed to evaluate the wheelchair's performance. A control weight optimization algorithm which can adapt the user's control weight to his or her control ability is proposed. Evaluation tests validated the algorithm's adaptability to the difference of different users' control abilities. Results of the elderly home tests show that this proposed method can reduce the difficulty of wheelchair maneuver and the seamless cooperation between human and machine makes the movements of a wheelchair more comfortable and safe. Future work will focus on providing assists under the user's higher level intention like going to a bedroom, parking by a desk. The challengeable part will be how to integrate global environment information into the definition of evaluation indices.


This work is partly supported by the National High Technology Research and Development Program of China under grant 2006AA040203, the Natural Science Foundation of China under grant 60775062 and 60934006, and the State Key Laboratory of Robotics and System (HIT).

Fig. 10 Clearance between wheelchair and obstacle.