Shared Control Navigation of TIAGo in a Partially Observable Environment using Artificial Potential Fields and User Intent Prediction

In recent years, the use of autonomous mobile robots with the goal of increasing the productivity in different domains, has been on the rise. The field of AI robotics has seen many advances since its emergence with the early robots that were built in an effort to facilitate some work for humans. The human-robot interaction can be categorized into three methods of shared-control, shared-autonomy and shared-intelligence. Starting from shared-control and moving towards shared-intelligence, the robot progressively acquires more participation in its own final action. Shared-control, which is also the focus of this thesis work, involves the robot being able to apply small modifications to the user inputs such as altering the steering angle due to the existence of an obstacle. In shared-autonomoy, there exist pre-defined behavioral procedures for the robot that get activated in very certain occasions and moreover, there is some a priori knowledge available about the environment which is provided to the robot. Shared-intelligence can be considered as the method in which the robot has the most active role in determining the final decision. It makes use of several policies which are essentially behavioral guidelines for the robot, each policy generates a probability grid which will be fused with every other policy. The final action of the robot will be based on this fusion, which is basically the action, that according to the policies, is considered as the most probable for that situation. Therefore, the work presented in this thesis explores the topic of shared control navigation of a mobile robot, namely TIAGo from PAL Robotics [1], in a partially observable environment, while the navigation is locally planned using artificial potential fields (APF) created around the objects and the system is assisted by predicting the most probable intention of the user. The robot through its laser is providing the source for the repellors and attractors, along with its pose info that will be subscribed to by the assistance system that takes as input the global goals and the user velocity commands as well. Shared navigation represents the node that based on the data coming from the repellors and attractors, compute the final APF velocity. This velocity can be combined with that coming from the user as a final control phase. It’s also worth noticing that there is no direct link between the user input and the robot and also, there is no direct link between the shared navigation node and the robot either. This suggests that any velocity has to first go through the filter of the final control, to be properly weighted before being published to the robot. The experiments were carried out both in a simulated environment and on the real robot. The behavioral results of the robot were almost similar in the two environments although we did not manage to dive deep enough in the real-world experiments. The APF showed promising results while the utilized assistance system was originally tested in an object manipulation context and still needs more fine-tuning for our case of navigation and this would thus be among the future work on this thesis along with providing more in-depth results from the real-world experiments.