Robust sensor fusion for finding HRI partners in a crowd

How Much Information Does a Robot Need? Exploring the Benefits of Increased Sensory Range in a Simulated Crowd Navigation Task

“…In particular, our study is missing a direct validation in hardware: primarily, there is no guarantee that the relationship between sensor range and task performance would follow the same curve as we find here. That said, other studies have shown that crowd navigation is possible using local information only [8], which at least strongly suggests that there would be limited additional benefits to making larger ranges accessible to sensors.…”

“…It is interesting to note, in this context, that in many robotics (and related) studies, similar problems are tackled with rather differing hardware set-ups. For crowd navigation, for example, even recent work in HRI uses a larger sensorised environment [7], but there are also examples of research using only on-board sensors (see, for example, [8]). This, in turn, raises the question of the added value of sensorising a certain environment compared to operating only based on local information.…”

“…For convenience, we call the degree to which a robot can perceive global information of the environment "sensing sophistication", and in practical terms, we treat this simply as a distance from the location of the robot. Low sensing sophistication thus corresponds to a robot that can only perceive its immediate surroundings, most likely using only on-board sensors (such as in [8]), whereas high sophistication would imply access to information that would often require additional sensors in the environment (such as in [7]). Note that the latter is not an infeasible proposition, even for large environments; in fact, much work in autonomous vehicles is concerned with solving navigation problems through cooperative systems that offer accurate information for large areas [9].…”

“…To address this last hurdle, we propose, in this paper, to use simulation environments to explore the trade-off between sensing sophistication and task performance. We choose to do so using a crowd navigation task, simply because previous research [7,8] suggests that it can be solved with both local and global information strategies. Our aim is a proof of concept; that is, we demonstrate the set-up for investigating this trade of in simulation environments but we are not aiming for a high-fidelity simulation, as a robot capable of navigating human crowds would, in practice, have to take into account human behaviour in its navigation strategy, which would require sophisticated models of such behaviour.…”

Neural network based reinforcement learning for audio–visual gaze control in human–robot interaction

“…We believe this could be useful in many real scenarios, such as a conversation between a companion robot and a group of persons, where the robot needs to learn to look at people, in order to behave properly. The reason for using multiple sources of information can be found in recent HRI research suggesting that no single sensor can reliably serve robust interaction [20]. Importantly, when it comes to the employment of several sensing modalities in complex social interactions, it becomes difficult to implement an optimal policy based on handcrafted rules that take into consideration all possible situations that may occur.…”