Meditation as an effective BCI training protocol for controlling virtual wheeled robots

Experiment design

The subjects ($N=30$) were assigned into a control group (non-meditators) and an experimental group (meditators). Both groups had 15 members. The subjects were required to fill out a questionnaire giving information about their previous experience with BCI, their level of experience with mindfulness meditation, and other additional information. The control group inclusion criteria was healthy individuals over 18 years of age. The control group exclusion criteria was individuals with previous experience with meditation and individuals with previous experience with BCI. The experimental group inclusion criteria was healthy individuals over 18 years of age with previous experience with mindfulness meditation. The experimental group exclusion criteria was individuals with previous experience with BCI. Participants were recruited from the city of Rio de Janeiro, Brazil by reaching out to students and professors from two prominent local universities. In addition to promoting the study within university classes, efforts were also made to raise awareness among meditation groups in the community. The participants ranged in age from 18 to 65 years and consisted of 18 females and 12 males.

Signal acquisition

The subject brain’s signals were gathered through EEG records from the scalp using the Emotiv EPOC+. The Emotiv EPOC+ is a low-cost, portable, and easy-to-configure device (Figure 1), that can be used without a conductive gel: instead of this, the electrodes may be moistened in saline water. Emotiv EPOC+ has wireless connectivity, 14 channels for signal collection (AF3, F7, F3, FC5, T7, P7, O1, O2, P8, T8, FC6, F4, F8, AF4) and 128hz or 256Hz sampling frequency. It is usually sold with a software Control Panel (CP) (Figure 2) and with a software development kit (SDK). Emotiv EPOC+ probably is not as robust as healthcare EEG equipment, however, due to its advantages, it was also used in other research.^16–18

Figure 1. Emotiv EPOC+ headset, electrodes, USB cable, saline, and Bluetooth USB stick.

Figure 2. Emotiv Control Panel command training software.

Training commands

Emotiv BCI require their users to initially train the commands that will be sent to the system afterwards. In this process, the user gives examples of his brain patterns, performing mental imagery tasks, so that the system adapts itself, through a calibration of its classifier algorithm. In our experiments, the commands were trained using the CP software (Figure 2). First, a baseline referred as Neutral state had to be configured. To do so, the subject was asked to relax, setting free the flow of thoughts, avoiding muscle movements (including facial expressions) for approximately 15 seconds. During this time, the system recorded the subject’s brain signals. Subsequently, the subject trained the commands to be used to control the virtual wheeled robot. To perform this training, the subject focused his thoughts on MI tasks, avoiding real muscle movements (including facial expressions) for approximately 8 seconds. If it was successful, the subject was able to observe an object moving on the CP. The CP also shows the Skill Rating, a numerical value ranging from 0 to 100 that indicates how fine the system recognized the subject’s mental patterns (the training accuracy). The MI tasks were right wrist extension and flexion, used to move the robot forward and feet dorsiflexion, used to spin the robot around itself. During this time the system recorded the subject brain signals.

Wheeled robot simulation

We used the Gazebo to simulate the Turtlebot 3 wheeled robot moving on a virtual environment (Figure 3). To control the robot, we used the Robot Operating System (ROS). ROS is a collection of tools, programming libraries, and software design patterns aiming to simplify the elaboration of complex and robust robotic behaviors on a wide variety of robotics’ architectures. We used Emotiv SDK and ROS to develop an interface (Figure 4) that can read subjects’ mental activities and convert them in robot actions. The Emotiv SDK comes with two predefined commands, Push and Pull, plus the Neutral state. We configured the Push command to move the robot forward, the Pull command to spin the robot around itself, and the Neutral state to stop the robot.

Figure 3. TurtleBot wheeled robot in the virtual environment.

Figure 4. Emotiv SDK and ROS were used together to develop an interface able to read subjects’ intentions and convert them in robot actions.

Hardware and software

Hardware: Emotiv EPOC+ and PC with 3 Ghz CPU 8Gb RAM. Standard onboard graphic accelerator chipset. Emotiv software: Xavier Control Panel v3.3 and Emotiv SDK community edition. Robot control and simulation software: ROS Kinetic Kame distribution and Gazebo3. Statistical software: python 3.9 and scipy v1.10.1.

Experiment steps

In the initial step of our experiments the subjects were asked to comfortably sit down in front of a computer screen. Then, the electrodes were moistened in saline water, placed on the subjects’ scalps, and the Neutral state was configured.

After this initial step, the subjects were asked to train commands using MI tasks. To balance performance and available time, they accomplished each command 10 times. The Emotiv’s manufacturer does not indicate an optimum number of times for training a state, although he advises that the more repetitions are performed the greater will be the system’ ability to recognize trained commands. Indeed, this makes sense since there is a classification model supporting the system. The more data are provided the better the model can generalize, as machine learning models demonstrated.

After the Push command training, subjects were asked to guide the robot in the simulated environment through the Emotiv EPOC+. Trained commands were then used to control the virtual wheeled robot.

In Test 1, the subjects tried to drive the robot to cross the virtual environment in a straight line, as fast as possible (Figure 5).

Figure 5. Test 1: Crossing the Gazebo simulated environment in a straight line.

In Test 2, the subjects also tried to drive the robot by crossing the virtual environment in a straight line, but this time they had to stop for 10 seconds at 3 predefined positions (Figure 6).

Figure 6. Test 2: Crossing the Gazebo simulated environment with stops at predefined positions.

In Test 3, the subjects had to train a second command used to move the robot through curves. For that purpose, the Pull command was trained in CP. Then, the subjects were instructed to move the robot to complete a rectangle shape circuit (Figure 7).

Figure 7. Test 3: Performing a rectangular move in the Gazebo simulated environment.

Training accuracy

The subjects trained Push and Pull commands. An unpaired t-test¹⁹ was done to compare the sample mean of the training accuracy values achieved by meditators and by non-meditators ($p<0.05$).

The result (Table 1) showed that meditators had greater training accuracy than non-meditators for both Push and Pull commands. The training accuracy sample mean for the Push shows to be greater than for the Pull command. Indeed, the Pull command was trained after the Push command, the classifier had to learn how to distinguish between the two commands. So, it was required that each subject is in two mental states, distinct from each other, thus considerably increasing the difficulty, especially for the command trained later.

Test 1 - Straight ahead moves

The subjects tried to traverse the virtual environment straight ahead as fast as possible. The sample mean time (in seconds) required to complete the task was smaller for meditators. An unpaired t-test¹⁹ shows that results (Tables 1 and 2) are statistically significant ($p<0.05$).

This result indicates that mindfulness meditation can effectively help BCI users, especially in tasks in which it is necessary to maintain focus for a long time. It also shows that mindfulness meditation is positively correlated to concentration skills. To perform the task of crossing the virtual environment straight ahead, a subject had to keep his focus on a single command, maintaining mental focus for as long as possible. Any distraction, like an external noise, or anxiety, arriving at the subject’s mind, may affect his brain waves and, consequently, hinder the recognition of the command by the system. Of course, this may compromise the remaining time to accomplish the task. Therefore, the mindfulness meditation training surely helped the subjects to avoid such distractions while using the BCI system to control the robot.

Test 2 - Stop and go moves

The subjects tried to traverse the virtual environment stopping at predefined places. The Barnard exact test²⁰ on a 2×2 contingency table (Table 3) showed that the results (Tables 1 and 2) are statistically significant ($p<0.05$).

These results indicate that mindfulness meditation can also assist BCI’s users in tasks in which it is necessary to switch between different mental tasks. In Test 2, the subjects had to use the Push command to move the robot forward, but also the Neutral state, to stop the virtual wheeled robot.

Test 3 - Rectangular shape moves

The subjects tried to control the virtual wheeled robot to move throughout a rectangular circuit in the virtual environment. The Barnard exact test²⁰ on a 2×2 contingency table (Table 4) showed that the results (Tables 1 and 2) are are statistically significant ($p<0.05$).

To complete a rectangular circuit in the virtual environment, the subjects had to control the virtual wheeled robot using two commands, one to move the robot forward and another to make curves. One of the main difficulties faced by the subjects during this test was to mentally alternate between these commands. Indeed, each command is associated with a specific mental pattern and swapping between them involves changing from one mental state to another. However, the subjects take some time to be able to immerse themselves into a particular mental state as they must reach a certain degree of concentration. Therefore, switching between commands to control the robot in real time was a big challenge for the subjects, leaving them, in some cases, to experiment anxiety, fatigue, and frustration, as they have reported in the questionnaires.

A possible explanation as to why the subjects belonging to the experimental group were better than the ones belonging to the control group is that the experience with meditation significantly increases the meditator’s attention control.²¹ This certainly has helped meditators to stay focused, avoiding external thoughts that would hinder the formation of the desired mental patterns.

Another suitable explanation is that the experimental group took some advantages from the mindfulness meditation experience to develop a more positive attitude when they were facing difficulties. Additionally, negative emotions due to anxiety, fatigue, frustration, etc. can decrease BCI performance because they can hinder the setup of stable mental states. The positive effects of meditation on emotion regulation were extensively discussed in our research.

A previous study¹¹ showed that subjects with no previous experience with meditation who underwent a meditation training, held once a week for 12 weeks, improved their accuracy by 10% on selecting letters on a computer screen using a MI BCI. Another study¹² with similar experiment setup showed that a short-term meditation training, held once a week for 4 weeks, produced improvements from 4.44% to 25.81% in subjects’ accuracy. In our study, we showed that experienced meditation practitioners were 144% more accurate than non-meditators controlling the virtual robot to traverse the virtual environment stopping at predefined places. We also showed that experienced meditation practitioners were 471% more accurate than non-meditators on controlling the virtual robot to move throughout a rectangular circuit in the virtual environment. Our findings indicate that the benefits that meditation brings to MI BCI users are more evident in complex tasks, as is the case of guiding a virtual robot using more than one command.

The advantages that meditation can bring in the execution of challenging tasks could also be observed in command training. We observed an improvement of the training accuracy for both MI commands among experienced meditation practitioners, when compared to non-meditators. Specifically, we found that meditation practitioners showed a 57% improvement in the training accuracy for the Push command, which was used to move the robot forward, and a substantial 76% improvement for the Pull command, which was used to spin the robot. The difference found between the commands can be explained by the fact that the Pull command was trained after the Push command. One should observe that once a command is trained, the second command is harder to train than the first one because the MI BCI system needs to distinguish between two distinct mental states. This difficulty arises due to the increased complexity of decoding and differentiating between multiple cognitive tasks.

Mindfulness meditation involves cultivating present-moment awareness and attention to one’s thoughts, bodily sensations, and emotions without judgment. Regular mindfulness practice can improve overall attentional control and focus. In the context of MI BCI, this enhanced attentional control could result in improved concentration and mental effort during motor imagery tasks, leading to more distinguishable EEG patterns and more accurate BCI control, as previously demonstrated.^9,10 Mindfulness meditation is known to reduce stress and anxiety levels. High-stress levels can interfere with EEG patterns and negatively affect BCI performance. By promoting relaxation and emotional regulation, mindfulness meditation could create a more conducive mental state for efficient MI BCI control.

Limitations

Our study is subject to certain limitations that should be considered. Despite implementing anonymous questionnaires, there remains a potential for participation bias that cannot be entirely eliminated. While participation was voluntary, the presence of self-selection bias may restrict the extrapolation of our findings. Additionally, the sample size in our study was relatively small, recruited from a limited geographic area, thereby raising concerns about the generalizability of the results. To enhance the validity of future research, it is crucial to include a larger and more diverse population, encompassing multiple cities or states. Furthermore, it is important to conduct tests involving individuals with motor disabilities, as they represent a significant target audience for BCI technology. In our study, subjects had a restricted time frame for training commands in the BCI system. Further investigation is recommended to verify whether similar outcomes can be achieved with extended training durations. Future works should also address raw EEG analysis seeing which electrodes showed the most difference between groups and for the different testing conditions.