*3.2. Wheelchair Movement to Reach Map-Based Desired Target*

The experiment was performed in an environment of 126.72 m2 which was divided into square grids of one map with a size of 8 × 11, in which each square has a size of 1.2 m × 1.2 m as shown in Figure 11. The wheelchair was installed to be able to move at the speed of 3 km/h for matching the processing speed of the system. An electrical wheelchair was installed with an RGB-D camera system and other equipment as shown in Figure 12. Information about the surrounding environment obtained from the camera system was processed by a computer and then transferred to the motor system of the wheelchair for motion control. In addition, in this research, we performed two experiments, including a self-control user and an automatic control user. In the self-control user model, the user can self-control commands such as going forward, backward, and turning right and left during the wheelchair movement. Meanwhile, the automatic control user mode means that the user can choose one of the targets by using EEG signals which are assigned to the targets to reach [36] and our proposed algorithm in the wheelchair control system is applied so that the wheelchair can automatically reach the chosen target.

Figure 13 shows the green real path of the wheelchair, which was controlled by the user during reaching the target. In particular, the discontinuous green path is the desired path in the real environment that the wheelchair needs to follow to reach the target, while the red path of the wheelchair is the path controlled by self-control mode using EEG signals [38] to go straight, turn left and right during reaching the destination. With the experiment using the self-control, the wheelchair moved according to the red path and then turned to the undesired direction shown by the red path and blue dash-dot ellipse. It means that in this case, the wheelchair could very easily have an obstacle collision. In addition, with the mode of the self-control, the movement of the wheelchair is unstable and discontinuous as shown in Figure 13. In particular, the wheelchair went straight, then stopped, turned right, and then was continuously interrupted during the movement time. It is obvious that the user was trying hard to control it to turn right or left and go straight.

**Figure 11.** The experimental environment. (**a**) The 1st view of the real environment; (**b**) the 2nd view of the real environment; (**c**) the 3rd view of the real environment; (**d**) the 2D grid map.

**Figure 12.** The wheelchair navigation system installed with devices.

**Figure 13.** The real path of the wheelchair movement and the reference path.

For improving the wheelchair control using the self-control mode, we used the proposed model with the semi-automatic control. With this mode, the user just needs to choose one typical target by using EEG commands and then the wheelchair will automatically move to reach the desired target with high stability and smoothness. In particular, using the environmental map in Figure 11a–c, the actual paths of the wheelchair after moving to reach the target were as shown in Figure 14b. Therefore, the moving process was re-calculated and the path positions of the wheelchair with the axes of X and Y were re-drawn for the purpose of the comparison with the simulation paths (blue arrows) as shown in Figure 14a. The starting point of the wheelchair is random and the wheelchair automatically determines its position on the map by identifying landmarks in the environment. In particular, in this case, the wheelchair determined it position on the grid map at the coordinate A(5,0) and the direction of the wheelchair *d* is Up. In the semi-automatic wheelchair, people with disabilities can control the wheelchair using EEG signals to select one of the commands on the interface screen with one sign corresponding to the target C(0,5). With the starting point A(5,0) and the target point C(0.5) selected, the RL model will produce a sequence of control commands for the path and then these commands are converted to control commands in the real environment for the wheelchair using Equations (8a)–(8d) as shown in Table 4.

**Figure 14.** Representation of the simulation route using the semi-automatic control and the wheelchair's real path (**a**) The blue arrow route simulated using DQNs; (**b**) the wheelchair movement path in the real environment using DQNs and the reference path.


**Table 4.** Wheelchair Control Commands Converted from Simulation Commands.

In addition, in this experiment, the actual path of the wheelchair with the proposed method of DQNs (blue path) is compared with the standard path (green dashed path), as shown in Figure 14b, for evaluating the wheelchair movement path and the simulated path. The results showed that the wheelchair could move to reach the desired target with the average error of ±0.2 m in the *X* axis and ±0.2 m in the *Y* axis.

The purpose of these experiments is to compare the results of the semi-automatic control using the RL method with the self-control by the user using the EEG signals. In particular, Figure 15a shows three graphs which represent the wheelchair movements, in which the blue path is that of the proposed mode and the red path is that of the self-control mode. From Figure 15a, it can be seen that the wheelchair's path when controlled by the semi-automatic control method is closer to the reference path than when using the self-control method. In addition, the wheelchair path using the semi-automatic control is smoother and more continuous than the path using the self-control. To clarify the two control methods, we recorded the wheelchair control commands during the movement to reach the destination.

**Figure 15.** The comparison of the stable movements of the wheelchair in two control methods (semiautomatic control and self-control). (**a**) The real paths of the two control methods and the reference path; (**b**) the control sequences of the two control methods.

In Figure 15b, the control commands are shown on the vertical axis with the values of −2, 0, 1, 2 corresponding to the commands to turn left, stop, go straight, and turn right. Therefore, it could be seen that the wheelchair moved with high stability in the case of the semi-automatic control with different movement environments compared to the mode of the self-control user. In addition, the result showed that the automatic control user mode spends less time on wheelchair movement with the average of about 80 s compared to that of the self-control user with the average time of about 95 s.

In another case, Figure 16d shows the simulation paths (blue arrows) of the wheelchair based on the environmental map in Figure 16a–c when the wheelchair moves from O(0,0) to C(0,5). From Figure 16e, it can be seen that the wheelchair's path controlled by the semiautomatic control method is shorter and smoother compared to the self-control method. Further, the semi-automatic control method has an average error of 0.1 m in the *X* axis and 0.3 m in the *Y* axis compared with ±0.5 m in the *X* axis and ±0.5 m in the *Y* axis of the self-control method. With Figure 16f, it can be seen that the wheelchair moved with high stability in the case of the semi-automatic control with different movement environments compared to the mode of the self-control user.

**Figure 16.** The comparison of the stable movements of the wheelchair in two control methods (semiautomatic control and self-control). (**a**) The 1st view of the real environment; (**b**) the 2nd view of the real environment; (**c**) the 3rd view of the real environment; (**d**) the blue arrow route simulated using DQNs; (**e**) the real paths of the two control methods and the reference path; (**f**) the control sequences of the two control methods.
