Deliverable 3.6: Shared control navigation

Work Package 3 of the CrowdBot project focuses on navigation. Half of our prototype crowdbots (Pepper and cuyBot) are designed to be fully autonomous and so the navigation algorithm must deal with both global and local aspects of planning. However, the other two crowdbots (smart wheelchair and Qolo) are designed to support human users improving their quality of life. In this case, the human’s intention is the high-level global navigation plan, but they may need assistance in planning and executing the local navigation at the operational level. Deliverable 3.6 focuses on the paradigm where the user’s continuous control input is safely blended with the robot’s local planner in a process we call Shared Control. Within Work Package 3 of the CrowdBot project, the UCL team has concentrated on refining two primary algorithms for shared control (probabilistic dynamic window approach and user intent modelling), both in terms of gathering data in real pedestrian environments and improving performance and user experience in simulation.
Probabilistic shared control using the dynamic window approach and generalised velocity obstacles (PSC-DWAGVO) was selected as an approach due to identified weaknesses in conventional DWA algorithms when it came to avoiding moving pedestrians. Combining the velocity obstacle paradigm with PSC allows the wheelchair to select a trajectory that is most likely to satisfy the user and simultaneously preserve safety, without the risk of freezing in place due to a rigid avoidance strategy. In order to support the development of this algorithm, a large, novel dataset of wheelchair-pedestrian interactions was collected, filling a vital gap in the available data. This dataset was used to create a simulated testing environment with realistic ‘pedestrian’ agents that allows the algorithm to be rapidly iterated and tested in the absence of human participants (this has become of even more crucial importance during the current pandemic).
Section 2 gives an overview of our custom built smart wheelchair platform and other shared control wheelchairs in the literature. Then, the motion planning and blending strategy is discussed in the Related Work Section 3. First, to give background context, the standard Dynamic Window Approach (DWA) motion planning algorithm is presented followed by the Generalised Velocity Obstacles (GVO) extension for environments with moving obstacles (crowds of people in our case). We then explore the baseline linear blending strategy and its extension to Probabilistic Shared Control (PSC), which formulates blending within a statistical framework and sets the tone for our approach.
Shared control navigation using Probabilistic Shared Control (PSC) is presented in Section 4. Our proposed model is a hierarchical framework for collision avoidance, which treats static and dynamic obstacles separately thus allowing for more flexibility and transparency in control, whilst improving computational performance. The first layer runs DWA to construct sets of Reachable Admissible Velocities (RAV), the second layer formulates Generalised Velocity Obstacles, and the final layer performs the actual blending using PSC, resulting in the final trajectory output.
The above-mentioned model is implemented and validated in a Unity3D ROS simulation environment containing the smart wheelchair platform with static and moving obstacles. All sensors’ characteristics apart from the RGBD camera are modelled. In our experimental results, we focus on three core metrics, which stem from Deliverable 1.3 namely: number of collisions, task completion time and “agreement” (between the user and the robot). Our results show that the proposed approach is a good step towards enabling a shared control wheelchair to navigate safely in a highly dynamic and crowded environment.
For 1D and 2D flow of crowds, in our simulated PAMELA facility, results indicate that in a crowd-robot navigation task, our shared control navigation strategy has a small, non-significant, effect on the crowds. This is expected as the current strategy does not aim to minimize such effects.