**3.3 Action game genre**

Most games (18/28) belong to action games whose players try to control the gaming character to present physical activity, such as jumping or moving to one side, to overcome the challenges and gain rewards. The following paragraphs introduce four main subgenres of action games (2D avoiding obstacles, 2D ball to basket, 3D Cybathlon brain-running, VR first-person action) and other games.

*2D avoiding obstacles* (2D AO) games (5/18) are a series of games where players use different imaginations to control the character's direction to avoid obstacles and collect rewards. In 2007, one study [33] presented a "Jump and Run" game whose ten healthy players used their adept imagination (right hand, left hand, or feet) to send jump commands to avoid the intermittent obstacles. Two out of five subjects presented a high accuracy of more than 90%. An 'arcade game' with similar up and down movement was present with an accuracy of 70% among ten subjects [34]. In 2011, a spaceship control game was provided [35]. Gamers used two out of four motor imageries (left hand, right hand, both feet, or tongue) to control the spaceship to move left or right to avoid falling asteroids. The synchronized game score also rises when the character passes asteroids successfully. Two out of three subjects reported increasing classification accuracy during this gameplay. A similar game is a design where a spaceship is replaced by a car, adding a coin reward [36]. Feet MI was linked with left movement, and right-hand MI could control the car to the right. An average multiple event accuracy (ME\_Acc) could reach 78.3% among four participants. One game [37] provided an environment where three classes of MI could be presented. Gamers could move to the left or right by ipsilateral hand imagery to collect coins. A jump command is sent by feet of imagination when the subject has to avoid the snake on the lane. After three gameplay sessions, an increasing average classification rate (53% to 63%) and game score (1600 to 1900) are presented among 14 BCI naïve healthy participants.

*2D ball to basket* (2D B2B) games (3/18) are games where players use one side of their body image to control a falling ball landing on the ipsilateral side of the basket. In 2003, one article indicated the feasibility of applying this game to 4 paraplegic patients [23]. 3 out of 4 participants succeeded in controlling the falling ball to an exact color basket by using two optimized MI from left hand MI, right hand MI and feet MI. One study in 2010 revealed the feasibility of using this type of game for stroke patients with a moderate classification accuracy of 60%-75% among five novices BCI subjects [2]. This result is similar to a previous study [38] with six healthy users (average classification accuracy of 69.2%). In 2016, nine healthy users were trained to use left or right-hand MI to control a platform (basket demo) to save a falling parachutist (ball demo) [31]. All the accuracy performances are higher than 70%.

*3D Cybathlon brain-running* (3D CyR) is a game (3/18) designed for the first Cybathlon competition held for a disabled cohort in 2016 in Switzerland [39]. Cybathlon competitors, called pilots, in BCI Race, present four different MIs depending on the areas the role needs to pass through. The result is ranked from the shortest runtime in order. In the MIRAGE team, whose pilot is a post-stroke patient, the average runtime could reduce from 178s to 143s during training and 196s (rank 11/11) in competition [40]. The optimized 2 MI classes are right hand MI, both feet MI, with other two mental subtraction, and auditory imagery. In the same competition, two tetraplegic pilots in the Brain Tweakers team [41] present 90s (rank 1/11) and 123s (rank 2/11) runtimes in the first race. In the final round, these two pilots present 125s (1/8) and 190 (4/8). Their control strategy is a 3-class MI (right hand, left hand, and both feet) combined with resting. Those two groups of pilots are training regularly for months before participating Cybathlon competition. In contrast, one research [42] shows that a healthy cohort trained with only two sessions could reach an accuracy of 68.62% with 2-classes of MI (right hand MI and stomach MI).

*VR first-person games* (VRFP, 4/18) is a group of action games with visual reality (VR) devices to present a first-person vision gamified feedback. From the players' view, no realistic but misleading interruption such as a passing-by researcher would occur so that players could fully be engaged in the game. In 2013 [43], 12 healthy subjects with one post-stroke patient accomplished a VR B2B game. Participants tried to use the ipsilateral hand MI to control the virtual hands to the right side for picking the falling basketball. Results show that although the disabled users only trained for one session, the final accuracy could reach 77%, whereas the highest accuracy among two-session healthy users is 70.67%. A similar VR B2B game designed in 2019 with 4-class MI (left hand MI, right hand MI, both feet MI, and resting) presents a mean accuracy of 70% with 10 participants [44]. One game with not only visual but also auditory and haptic feedback, called "NeuRow," was present in 2016 [45]. 13 healthy users use their left- or right-hand MI to navigate the boat towards the same side. Two high-fidelity arms enhance the MI of users by linking the imagery environment with the real world. The classification score (70.7%) in NeuRow compared with the other six studies with the same classification method indicates that the NeuRow has scored the best. Another VR MI-BCI game designed in 2019 is a simplified shooting game [46]. Nineteen users try to destroy the asteroid (VR DA) with the same side hand MI. After receiving the MI commend, the embodied visual hand would move to switch on the trigger of the weapon to shoot the asteroid. The average peak accuracy among these users could reach 75.84%.

In 2003, one 3D first-person shooting game (3DShT, 1/18) was tested among four healthy participants [24]. Since it is the earliest MI-BCI game, researchers did not limit the users to control the game in a certain fixed MI way. Instead, participants are
