A local exploration cost function for removing alignment constraints in perception-aware path following

Kurzfassung

The research field of space exploration is growing and plays a crucial role in replacing human presence, as manned missions to Mars are dangerous and time-consuming.The Deutsche Luft- und Raumfahrt RMC research group is tasked with building aprototype of a mobile Lightweight Rover Unit (LRU) for autonomous navigation, ascommunication with Earth would result in an unbearable response delay during re-motely controlled exploration. At the current state, the rover is able to find a path to apredefined target and reach it safely using visual perception from a pan-tilt camera todetect obstacles along the path, and recalculating the path if necessary. To do so, thecamera always follows the path ahead, thereby ignoring to collect important informa-tion of the environment alongside the path. The goal of this study is to extend visual exploration of the LRU’s environment to collect more data about the Martian surface. Our approach incorporates additional degree of freedom of the camera in controlling the pan angle while following a predefined path. The challenge is to maintain sufficient focus on the path to avoid collisions or traveling through unknown terrain in autonomous navigation, while allowing enough freedom for visual exploration. Recent work in autonomous exploration has focused on finding the most efficient path to cover all the unknown surroundings, given the current knowledge of the map, to explore the entire environment. However, our work differs from conventional approaches, since the target position in our scenarios is predefined. Therefore, our objective is to extend the visual exploration along the path towards the specified goal. To find the best pan orientation, we introduce a simplified simulation of the visual perception and traversal through the world map. In order to evaluate and compare different approaches we examine them for the following metrics: (1) whether a collision occurred or the rover passed through unseen terrain, (2) the length of the path traveled, (3) the total exploration coverage along the path, (4) and the time required to determine the pan orientation. We sample over candidate pan angles and compare the potential seen costs, defined here as inverse reward function from information gain that the camera would capture in the travel environment. We aim to identify the best cost function for selecting best-suited pan orientation from the following: (1) static pan orientation aligned to the rover’s body, which does not utilize the full range of pan motion of the camera and therefore not necessitate in a cost function; (2) a cost function resulting in camera following the path ahead; (3) a cost function, inspired by Boyu Zho et al. [1], incorporating the lateral distance to the path in combination with the longitudinal distance relative to the rover’s current position; (4) and a cost function prioritizing the closest distance to the boundary of unknown areas. We show that the proposed cost function based on previous work of Boyu Zho et al. [1] is the best-suited strategy to allow visual exploration of the environment while ensuring a safe travel. [1] B. Z. et al. “RAPTOR: Robust and Perception-Aware Trajectory Replanning for Quadrotor Fast Flight.” In: IEEE Transactions on Robotics 37.6 (2021)