Bhaskar Trivedi

In the real world, humans often collaborate with others without direct communication. To do this successfully, they have to infer their intentions and choose actions that complement the predicted actions of their collaborators to perform the task efficiently. Since the peer’s state and action are generally not directly observable, these are usually estimated based on environmental change and then used to predict the intention. While humans can achieve this easily, this form of collaboration is difficult for artificial intelligent agents operating in partially observable environments, leading to agent architectures that do not attempt to explicitly infer other agents’ intentions but rather rely on additional knowledge or reactive collaboration, relying on the steadystate character of other agents.In this paper, we propose an agent model that explicitly defines and utilizes estimates of other agents’ intentions to yield more effective collaboration in decentralized partially observable domains, where each agent’s knowledge of and current belief state in the environment can be different. The resulting agents explicitly estimate other agents’ intentions from their observationsand utilize these estimates in a Reinforcement Learning process on a modified Dec-POMDP model to learn collaborative strategies. Initial experiments in a simple, partially observable collaborative manipulation domain show the ability of these intention-aware agents to learn optimal hierarchical strategies faster and more stably than equivalent agents without intention awareness.

Unmanned Aerial Systems (UAS) have become increasingly popular and have been identified as a good platform for a range of tasks from surveillance and inspection to delivery and maintenance. In many of these applications these systems have to operate in environments that are frequented by people or that contain sensitive infrastructure and in which the operation of UAS thus poses physical risk in terms of damage in case of vehicle failure or psychological or privacy risks which would make their operation less acceptable. To increase the use of these systems it is thus important that they can take into account these risks when determining navigation strategies. While this can sometimes be done based on prior information, such as street and building plans in cities, a priori information is often not complete, making it essential that risk representations can be augmented in real time based on sensor information. This paper presents an approach to risk map augmentation that uses learned risk identification from aerial pictures to fuse additional information with prior data into a dynamically changing risk map that allows effective re-planning of navigation strategies.