CamerAI is an autonomous chase camera that uses a PPO-trained neural network. Our results suggest that a simple, fully connected network with only two hidden layers of 128 neurons can perform basic chase camera functionality. We contribute a set of inputs and outputs with their respective coordinate spaces, as well as a custom reward function that can be used to train the network with reinforcement learning. Our findings highlight the importance of correct coordinate spaces, the need for a continuous reward function, normalization as well as timely resets during training to allow the network to explore its environment. We additionally present an evaluation of the output of CamerAI during 10 hours of chasing a bot that is randomly exploring a commercial video game map in which CamerAI was able to keep the player visible 96 % of the time.

The coupling between perception and action has seldom been explored in sophisticated motor behaviour such as 3D pointing. In this study, we investigated how 3D pointing accuracy, measured by a depth estimation task, could be affected by the target appearing in different visual eccentricities. Specifically, we manipulated the visual eccentricity of the target and its depth in virtual reality. Participants wore a head-mounted-display with an integrated eye-tracker and docked a cursor into a target. We adopted a within-participants factorial design with three variables. The first variable is Eccentricity: the location of the target on one of five horizontal eccentricities (left far periphery, left near periphery, foveal, right near periphery and right far periphery). The second variable is Depth at three levels and the third variable is Feedback Loop with two levels: open/closed. Eccentricity is refactored into Motion Correspondence between the starting location of the cursor and the target location with four levels: periphery to fovea, fovea to periphery, periphery to periphery, fovea to fovea. The results showed that the pointing accuracy is modulated mainly by the target locations rather than the initial locations of the effector (hand). Visible feedback during pointing improved performance.

Bipedal walking is generally considered to be the most natural and common locomotion technique in the physical world, for humans, and the most presence-enhancing form of locomotion in virtual reality (VR). However, there are significant differences in the way people walk in VR compared to their walking behaviour in the real world. For instance, previous studies have shown a significant decrease of gait parameters, in particular, velocity and step length in the virtual environment (VE). However, those studies have only considered short periods of walking. In contrast, many VR applications involve extended exposures to the VE and often include additional cognitive tasks such as way-finding. Hence, it remains an open question whether velocity during VR walking will further slowdown over time or if users of VR will eventually speed-up and adapt their velocity to the VE and move with the same speed as in the real world.

We report an experiment in which participants compared the size of a visual sphere to a haptic sphere in VR. The sizes from two modalities were either congruent or conflicting (with different disparities). Importantly, three standard references (small, medium and large) for haptic sizes were used with the method of constant stimuli. Results show a dominant functional priority of the visual size perception. Moreover, observers demonstrated a central tendency effect: over-estimation for smaller haptic sizes but under-estimation for larger haptic sizes. The results are discussed in the framework of adaptation level theory for haptic size reference, and provide important implications for the design of 3D visuo-haptic human-computer interaction.

We report on findings from two focus groups for designing an application for balance training at home with an augmented reality virtual coach. Following a User-Centered Design approach, we performed two focus groups with older adults at the early stages of development. Focus group participants were shown a prototype using a Meta 2 head mounted display. Their movements were tracked using a Kinect 2. The virtual coach gave balance training instructions and demonstrated their correct performance. Results suggest that, given the trade-offs of traditional health care, older adults are positive towards using an AR coach for their balance training.

We evaluate the ability of locally present participants to localize an avatar head's gaze direction in 360° hosted telepresence. We performed a controlled user study to test two potential solutions to indicate a remote user's gaze. We analyze the influence of the user's distance to the avatar and display technique on localization accuracy. Our experimental results suggest that all these factors have a significant effect on the localization accuracy with varying effect sizes.

Interaction tasks in virtual reality (VR) such as three-dimensional (3D) selection or manipulation of objects often suffer from reduced performance due to missing or different feedback provided by VR systems than during corresponding realworld interactions. Vibrotactile and auditory feedback have been suggested as additional perceptual cues complementing the visual channel to improve interaction in VR. However, it has rarely been shown that multimodal feedback improves performance or reduces errors during 3D object selection. Only little research has been conducted in the area of proximity-based multimodal feedback, in which stimulus intensities depend on spatiotemporal relations between input device and the virtual target object. In this paper, we analyzed the effects of unimodal and bimodal feedback provided through the visual, auditory and tactile modalities, while users perform 3D object selections in VEs, by comparing both binary and continuous proximity-based feedback. We conducted a Fitts' Law experiment and evaluated the different feedback approaches. The results show that the feedback types affect ballistic and correction phases of the selection movement, and significantly influence the user performance.

We introduce Camera Time Warp (CamWarp), a novel reprojection technique for video-see-through augmented reality, which reduces the registration error between captured real-world videos and rendered virtual images. Instead of rendering the image plane locked to the virtual camera, CamWarp renders the image plane at the real-world position it was captured at, and compensates for potential artifacts. We conducted two experiments to evaluate the effectiveness of CamWarp. In the first experiment participants were asked to report subjective discomfort while moving their head in a pattern inspired by the ISO 9241-9 Fitts' Law task at different speeds while the video feed was rendered at varying frame rates. The results show that the technique can significantly reduce subjective levels of discomfort and cybersickness symptoms for all tested configurations. In the second experiment participants were asked to move physical objects on a projected path as quickly and precisely as possible. Results show a positive effect of CamWarp on speed and accuracy.

Telepresence systems have the potential to overcome limits and distance constraints of the real-world by enabling people to remotely visit and interact with each other. However, current telepresence systems usually lack natural ways of supporting interaction and exploration of remote environments (REs). In particular, single webcams for capturing the RE provide only a limited illusion of spatial presence, and movement control of mobile platforms in today's telepresence systems are often restricted to simple interaction devices. One of the main challenges of telepresence systems is to allow users to explore a RE in an immersive, intuitive and natural way, e.g., by real walking in the user's local environment (LE), and thus controlling motions of the robot platform in the RE. However, the LE in which the user's motions are tracked usually provides a much smaller interaction space than the RE. In this context, redirected walking (RDW) is a very suitable approach to solve this problem. However, so far there is no previous work, which explored if and how RDW can be used in video-based 360° telepresence systems. In this article, we conducted two psychophysical experiments in which we have quantified how much humans can be unknowingly redirected on virtual paths in the RE, which are different from the physical paths that they actually walk in the LE. Experiment 1 introduces a discrimination task between local and remote translations, and in Experiment 2 we analyzed the discrimination between local and remote rotations. In Experiment 1 participants performed straightforward translations in the LE that were mapped to straightforward translations in the RE shown as 360° videos, which were manipulated by different gains. Then, participants had to estimate if the remotely perceived translation was faster or slower than the actual physically performed translation. Similarly, in Experiment 2 participants performed rotations in the LE that were mapped to the virtual rotations in a 360° video-based RE to which we applied different gains. Again, participants had to estimate whether the remotely perceived rotation was smaller or larger than the actual physically performed rotation. Our results show that participants are not able to reliably discriminate the difference between physical motion in the LE and the virtual motion from the 360° video RE when virtual translations are down-scaled by 5.8% and up-scaled by 9.7%, and virtual rotations are about 12.3% less or 9.2% more than the corresponding physical rotations in the LE.

Conventional input devices such as the mouse and keyboard lack in intuitiveness when it comes to 3D manipulation tasks. In this paper, we explore the use of accelerometer and digital compass equipped mobile phones as 3-DOF controllers in a 3D object rotation task. We put the mobile phone up against the established standards, a mouse and a touch pen and compare their performance in a rotation task. Our preliminary evaluation indicates that for this type of task, with only 5 minutes of practice the mobile phone is significantly faster than both the mouse and the touch pen.