Description of the candidate: We are looking for the perfect PhD candidate to find out how to enable robots to interact with humans in the wild considering the perceptual and computational limits they have. The candidate will have the chance to explore practical machine learning and more formal methods to develop the AI controller of social robots. There will also be the opportunity for interdisciplinary collaboration to look at how humans and other organisms solve similar problems.
It will be crucial to be passionate about ideas and challenges.
Applicants are expected to have good programming skills and be interested in further improving them.
Knowledge of statistics, control systems theory, artificial intelligence, computer vision, as well as machine learning methodologies, and libraries would be an important plus. Similarly, the ability to understand and design psychological tasks as well as use statistical methods to evaluate experimental results and human-robot interaction effectiveness would be valuable. Experience with real-time 3d engines and/or VR platforms, such as Unity3D, Unreal and similar, or with robotic platforms will also be considered positively.
Description of the field:
In the last 10 years, with the advent of modern deep learning methodologies, substantial performance improvement has been observed in perception for robots and other artificial systems. However, interactions with unstructured environments pose high challenges due to the variety of conditions and crucial sensory limits, such as occlusions and limited FOV. This position will focus on the study and development of systems that can perceive others’ states in unstructured environments and predict their actions, intentions and beliefs.
A possible line of research would focus on adaptive and social active perception mechanisms that enable to dynamically deal with sensory limits and have received limited attention but play a crucial role in human perception (Ognibene & Demiris, 2013, Lee, Ognibene et al. 2015). It has been recently shown that such mechanisms may substantially improve learning performance other than execution efficiency and even enable online adaptation to new environments [Ognibene & Baldassarre, 2015], however, these properties have not been fully scaled to social conditions yet. Moreover, active perception also plays a crucial role also when interacting with other agents who add relevant scene dynamics and may occlude important information. At the same time agents may have their own sensory limits and active perception strategies that must be scrupulously parsed to support effective social interaction [Ognibene, Mirante et al, 2019], e.g. false beliefs and theory of mind [Bianco & Ognibene 2020]. Most importantly, social interaction increases the demand for integration of information about task and context, i.e. simultaneous perception of the states of other agents, their effectors and other scene elements which can be strongly affected by the limited field of view and challenging for active perception due to the necessity to focus on the right element at the right time [Ognibene, Chinellato, et al 2013] and adapt to different types of interaction. The work may not only focus on advancing technical performance but on understanding and modeling how humans perform and adapt social perception or on how to design active social perception to improve the perceived quality of human-robot interactions.
Bianco, F., & Ognibene, D. (2020, March). From psychological intention recognition theories to adaptive theory of mind for robots: Computational models. In Companion of the 2020 ACM/IEEE International Conference on Human-Robot Interaction (pp. 136-138).
Ognibene, D., Mirante, L., & Marchegiani, L. (2019, November). Proactive intention recognition for joint human-robot search and rescue missions through Monte-Carlo planning in POMDP environments. In International Conference on Social Robotics (pp. 332-343). Springer, Cham.
Lee, K., Ognibene, D., Chang, H. J., Kim, T. K., & Demiris, Y. (2015). Stare: Spatio-temporal attention relocation for multiple structured activities detection. IEEE Transactions on Image Processing, 24(12), 5916-5927.
Ognibene, D., Chinellato, E., Sarabia, M., & Demiris, Y. (2013). Contextual action recognition and target localization with an active allocation of attention on a humanoid robot. Bioinspiration & biomimetics, 8(3), 035002.
Ognibene, D., & Demiris, Y. (2013). Towards active event perception. In Proceedings of the 23rd International Joint Conference of Artificial Intelligence (IJCAI 2013).