GONet++: Traversability Estimation via Dynamic Scene View Synthesis

Robots that interact with a dynamic environment, such as social robots and autonomous vehicles must be able to navigate through space safely to avoid injury or damage. Hence, effective identification of traversable and non-traversable spaces is crucial for mobile robot operation. In this paper, we propose to tackle the traversability estimation problem using a dynamic view synthesis based framework. View synthesis technique provides the ability to predict the traversability of a wide area around the robot. Unlike traditional view synthesis methods, our method can model dynamic obstacles such as humans, and predict where they would go in the future, allowing the robot to plan in advance and avoid potential collisions with dynamic obstacles. Our method GONet++ is built upon GONet. GONet is only able to predict the traversability of the space right in front of the robot. However, our method applies GONet on the predicted future frames to estimate the traversability of multiple locations around the robot. We demonstrate that our method both quantitatively and qualitatively outperforms the baseline methods in both view synthesis and traversability estimation tasks.