RRT*-GBO for Perception-aware Trajectory Optimization using NASA's Astrobee Robot

Kurzfassung

Robots can be used for on-orbit tasks in outer space, such as interacting with satellites. A free-flying robot such as NASA’s Astrobee should be able to autonomously navigate in space, but a common issue is that the robot’s vision-based localization system fails if its camera does not see enough visually descriptive landmarks. That makes the robot lose track of its position and orientation, and consequently fail to execute a desired trajectory. To avoid this scenario, this thesis presents a trajectory planner that guarantees the visibility of descriptive features while maintaining energy efficiency and satisfying motion constraints from the robot’s actuators and obstacles in the environment. The visibility consideration in the planning of a trajectory is referred to as perception-aware planning. The trajectories are optimized by exploring all six translation and rotation degrees of freedom of the SE(3) domain with the RRT*-GBO algorithm, which uses an effective sampling-based approach to find initial guesses that converge to globally optimal solutions. The time efficiency of the RRT*-GBO algorithm is highlighted in different scenarios. Additional time savings are obtained with an offline precomputation of the feature visibility metric for a given environment. Our solution is experimentally validated for a rendezvous task between two Astrobee robots in the International Space Station (ISS) and for a satellite capture task on DLR’s On-Orbit Servicing Simulator (OOS-SIM). These two experiments demonstrate a significant improvement in the accuracy of the robot’s localization system by planning perception-aware trajectories.