Topological mapping for traversability-aware long-range navigation in off-road terrain

TL;DR

A novel traversability analysis method is demonstrated, predicting the existence of a safe path towards a target in an image, and an exploration planner is presented, efficiently covering an unknown off-road area with unknown traversability using a frontiers-based approach.

Abstract

Autonomous robots navigating in off-road terrain like forests open new opportunities for automation. While off-road navigation has been studied, existing work often relies on clearly delineated pathways. We present a method allowing for long-range planning, exploration and low-level control in unknown off-trail forest terrain, using vision and GPS only. We represent outdoor terrain with a topological map, which is a set of panoramic snapshots connected with edges containing traversability information. A novel traversability analysis method is demonstrated, predicting the existence of a safe path towards a target in an image. Navigating between nodes is done using goal-conditioned behavior cloning, leveraging the power of a pretrained vision transformer. An exploration planner is presented, efficiently covering an unknown off-road area with unknown traversability using a frontiers-based approach. The approach is successfully deployed to autonomously explore two 400 meters squared forest sites unseen during training, in difficult conditions for navigation.

Figures (5)

• Figure 1: Our topological map allows for a sparse representation of the environments, with nodes being GPS waypoints and edges representing reachability between neighboring nodes, from a traversability point of view. In (a), we first initialize the map by computing frontiers and if the path towards them is traversable (\ref{['sec:trav']}). The closest reachable frontier is selected, and the BC controller goes to its position (\ref{['sec:bc']}). In (b), after the robot has reached its target node, we recompute the frontiers and select the next one. This process is repeated until no reachable frontier nodes are left in the desired zone of exploration, as shown in (c).
• Figure 2: Here the traversability labeling user interface is shown. On the left, we have the two panoramas taken from the two different positions. When one node is visible from the other, its position is projected in the panorama of the other node. On the right, the user has a more global view of the two nodes' position, helping relate the nodes' position to each other. In this example, the yellow node is visible from the blue node. Here there are no visible obstacles or other visible problems preventing the robot from reaching the yellow node from the blue node. This example is thus labeled as traversable by the labeller.
• Figure 3: An illustration of the frontier creation algorithm. (a) starting map (b) creation of the occupancy grid (c) computing the frontier cells in the grid (d) creating the frontier nodes from the cells and adding them to the map (e) final result. The graph, containing traversability information, can then be used to plan further exploration.
• Figure 4: Confusion matrices for the training set, as well as manually labeled data from the two test runs. The rows represent the ground truth, while the columns represent the model's prediction.
• Figure 5: Topological maps generated using our autonomous exploration process on two unseen environment. Nodes in the graph are color-coded as a function of the terrain height, highlighting the rough and hilly nature of the terrain. In testing zone A, (a) demonstrate how the robot found a path in between a large tree and a pile of branches that would have been difficult to traverse. In (b) another pile of branches is perceived as both untraversable and traversable, highlighting the viewpoint sensitivity of our approach. Indeed, there is a large branch high off the ground which could block the robot depending how the robot is approaching the pile. Now in testing zone B, (c) shows multiple fallen tree completely block a zone from being explored. However, in (d), we see a large fallen tree correctly detected as untraversable, which the robot is able to find a path around and keep exploring.