Active perception based on 2D LiDAR for social navigation in crowded environments

In the domain of social robotics, where robots dynamically interact with humans, the fusion of laser data with vision systems and Machine Learning (ML) techniques is crucial for navigation toward target people. Active perception not only reduces uncertainties but also enhances the sociability of robots by granting them a more profound comprehension of the social context in which they operate. In this thesis new active perception methods for social navigation in crowded environments are proposed that are based on 2D Light Detection and Ranging (LiDAR) in order to be used in public space by respecting privacy and without requiring additional sensors. The proposed methods consider a policy that measures and exploits the level of uncertainty in the people detection to choose the most useful poses for the robot to confirm the detection or to deny a previous wrong detection. The other two methods concern learning this policy based on a Deep Neural Network (DNN) and learning via Reinforcement Learning (RL) the most appropriate robot’s motions, to increase the certainty about the people in the environment. Results about the policy behavior show that this approach is feasible to be used on a real robot than the other two approaches that were tested only in simulations. Furthermore, the experiments at the IAS-Lab on the Take It And Go (TIAGo)++ revealed that the robot can adapt its behavior to the working scenario, in which it can accomplish its tasks successfully when deployed in challenging situations, that is the goal of active perception.