Geometric and Learning Methods for Robots to Navigate in Human Crowds with Application to Smart Mobility Devices
The thesis at hand is concerned with robots' navigation in human crowds. Specifically, methods are developed for planning a mobile robot's local motion between pedestrians, and they are evaluated in experiments where a robot interacts with real pedestrians as well as in simulations of a crowd and a robot. The thesis is divided in three main contributions.
The first contribution describes a novel method for non-holonomic robots of convex shape to avoid imminent collisions with moving obstacles. The method assists navigation by correcting steering from the robot's path planner or driver. Its performance is evaluated using a custom simulator, which replicates real crowd movements from a campus dataset, and corresponding metrics which quantify agents' efficiency, the robot's impact on the crowd, and the number of collisions. Further, the method is implemented and evaluated on the standing wheelchair Qolo. In the experiments performed, it drives in autonomous mode using on-board sensing (LiDAR, RGB-D camera and a system to track pedestrians). It avoids collisions with up to five pedestrians and passes through a door.
The second contribution studies the Acceleration Obstacle (AO) for enabling a robot's navigation in human crowds. The AO's geometric properties are analyzed and a direct sampling-free algorithm is proposed to approximate its boundary by linear constraints. The resulting controller is formulated as a quadratic program and evaluated in interaction with simulated bi-directional crowd flow in a corridor. A comparison to alternative robotic controllers is carried out, considering the robot's and the crowd's performance and the robot's behavior with respect to emergent lanes. Results indicate that the robot can achieve higher efficiency outside lanes.
In the third contribution, the problem for a mobile robot to navigate seamlessly in a human crowd is treated by an inverse reinforcement learning (IRL) approach. A novel feature is proposed to model costs of anticipated collisions between agents. The feature approximates agents' pairwise interaction energy, a function which prior work has derived empirically from crowd data as an interaction potential driving pedestrians' mutual avoidance. Using a recent framework to perform IRL from locally optimal examples in continuous space, cost functions which incorporate the novel feature are learned efficiently from high-dimensional examples of real crowd motion. Examples are obtained from two public datasets containing pedestrians' and wheelchair users' trajectories.
The learned models are evaluated and compared in how accurately their local optima model the training examples and test examples. Furthermore, predictions based on test examples' initial states only are generated similarly by optimization, and their distance to recorded ground truth is measured. Both models' predictions compare favorably to a recent related approach from the literature.
Finally, a control system which computes and executes in real-time an optimal trajectory according to the learned cost functions is implemented on a robotic wheelchair, to steer it between pedestrians perceived by an on-board tracking system. The robot is deployed on campus, where the controller's performance is evaluated qualitatively. Results show that the approach often generates apt motion plans, which complement pedestrians' motion in an efficient manner, albeit oscillations between locally optimal solutions may occur.
EPFL_TH10034.pdf
n/a
openaccess
copyright
15.24 MB
Adobe PDF
f6d0b19253033af6df614d293488d33c