Autonomous navigation in dynamic uncertain environment using probabilistic models of perception and collision risk prediction.

TL;DR: A novel probabilistic extension of the well known Rapidly-exploring Random Trees framework is proposed and this search algorithm is integrated into a partial planning anytime approach, updating on-line the decisions of the robot with the last observations.

Abstract: In this document we address the problem of autonomous navigation in dynamic unknown environ- ments. The key of the problem is to guarantee safety for all the agents moving in the space (people, vehicles and the robot itself) while moving toward a predefined goal. In contrast with static or controlled environments, high dynamic environments present many difficult issues: the detection and tracking of moving obstacles, the prediction of the future state of the world and the on-line motion planning and navigation. We moved our research starting from the fact that the decision about motion must take into account the limits of sensor perception, the velocity and future trajectory of the moving obstacles and that it must be related with the on-line updating of the world perception. Our approach is based on the use of probabilistic frameworks to represent the uncertainty about the static environment and the dynamic obstacles and we have developed decision methods that take explicitly into account such information for the autonomous navigation task. At first we focused our attention on reactive approaches. The developed approach is based on the integration between a probabilistic extension of the Velocity Obstacle framework within a dynamic occupancy grid (Bayesian Occupancy Filter). The dynamic occupancy grid gives a probabilistic and explicit representation of the perception limits and of the model error and noise while the Probabilistic Velocity Obstacles compute a probability of collision in time. The probability of collision and hence the choice of the next control, take into account limited sensor range, occlusions, obstacles shape and velocity estimation errors. Simulation results show the robot adapting its behaviour to different conditions of perception, avoiding static and moving obstacles and moving toward the goal. In the second part of the document we move toward more complex strategies, integrating a Partial Motion Planning algorithm with probabilistic representation of the environment and collision risk prediction. We propose a novel probabilistic extension of the well known Rapidly-exploring Random Trees framework and we integrate this search algorithm into a partial planning anytime approach, updating on-line the decisions of the robot with the last observations. We consider at first the case of a short-term prediction based on a classic Multi-Target-Tracking algorithm. Secondly, we consider the case of a longer term prediction based on a priori known typical patterns. We compare two kinds of patterns representation: the Hidden Markov Model representation and the continuous Gaussian Processes representation. The search and planning algorithm is adapted to these different kinds prediction and simulation results for navigation among multiple moving obstacles are shown.