Safe-Nav: learning to prevent PointGoal navigation failure in unknown environments

Abstract

Training robots to safely navigate (Safe-Nav) in uncertain complex environments using the RGB-D sensor is quite challenging as it involves the performance of different tasks such as obstacle avoidance, optimal path planning, and control. Traditional navigation approaches cannot generate suitable paths which guarantee enough visible features. Recent learning-based methods are still not mature enough due to their proneness to collisions and prohibitive computational cost. This paper focuses on generating safe trajectories to the desired goal while avoiding collisions and tracking failure in unknown complex environments. We present Safe-Nav, a hierarchical framework composed of the visual simultaneous localization and mapping (SLAM) module, the global planner module and the local planner module. The visual SLAM module generates the navigation map and the robot pose. The global planner module plans a local waypoint on the real-time navigation map. In the local planner module, a deep-reinforcement-learning-based (DRL-based) policy is presented for taking safe actions towards local waypoints. Our DRL-based policy can learn different navigation skills (e.g., avoiding collisions and avoiding tracking failure) through specialized modes without any supervisory signals when the PointGoal-navigation-specied reward is provided. We have demonstrated the performance of our proposed Safe-Nav in the Habitat simulation environment. Our approach outperforms the recent learning-based method and conventional navigation approach with relative improvements of over 205% (0.55 vs. 0.18) and 139% (0.55 vs. 0.23) in the success rate, respectively.