One step back, two steps forward: learning moves to recover from SLAM tracking failures

Abstract

Tracking failure in state-of-the-art visual SLAM has been reported to be frequent and hampers real-world deployment of SLAM solutions. Very recently, efforts have been made to avoid tracking failure using various methods (e.g. using deep reinforcement learning to plan a path where the risk of tracking failure is minimal). The results of those approaches are encouraging but are far from producing a failure-free visual system. Failure is inevitable in vision-based systems. Therefore, developing recovery (post-failure) solutions might be a better approach in developing more reliable systems. To this end, we propose a novel, easily trainable, deep recovery maneuver algorithm. Instead of predicting tracking failures, our algorithm predicts the back stepping move, post failure, which helps regain tracking. With our proposed simple approach, state-of-the-art SLAM manages to complete 13 out of 50 navigation episodes (13x improvement over the baseline). Furthermore, even for cases where our algorithm fails to complete the route, it traverses 2x longer distances as compared to baseline methods.