Traversability Estimation

Judging which parts of the terrain ahead can be crossed, which can’t, and how costly each path would be.

This is a question wheeled and legged mobile robots must answer before navigating: can each patch of ground ahead be crossed safely, and how difficult would it be? Traditional methods compute slope, roughness, and step height from an elevation map (a 2D map recording the ground height at every grid cell); this purely geometric approach mistakes tall grass or branches for obstacles, which is why later work added semantic information (grass, mud, water) and self-supervised methods that learn from the robot’s own walking experience. The result is usually output as a traversability map or costmap, handed to a path planner to choose a route. Wild Visual Navigation, from ETH Zurich and Oxford in 2023, let the quadruped ANYmal learn to tell walkable vegetation apart from real obstacles after less than 5 minutes of on-site training.

ExampleUsing Wild Visual Navigation, the ANYmal quadruped, after being briefly led by a person, can push through tall grass and go around tree trunks on its own, completing a 1.4 km trail navigation.

Also called

Traversable-Area Analysis, Terrain Traversability

Related

Elevation Map · Costmap · Perceptive Locomotion · Rough-terrain Locomotion · Navigation · Semantic Segmentation

Sources

Fast Traversability Estimation for Wild Visual Navigation (RSS 2023, arXiv:2305.08510)