Autonomous off-road driving at speed is hard because the vehicle's motion depends heavily on local terrain, yet many learned dynamics models treat perception as a simple, unstructured input. The authors of SpatialBEV argue that this ignores the inherently spatial nature of the problem: which patch of ground the robot is about to cross matters as much as what that ground looks like.
Their approach queries bird's-eye-view features at the predicted state at each step, feeding current geometric and semantic cues into the next prediction. They also introduce a benchmarking method that automatically mines a dataset for its most difficult scenarios, making evaluation more targeted than random splits.
Across datasets collected in different seasons, SpatialBEV reduced prediction error by 17.1% relative to physics-based and learned baselines. On hardware at speeds of at least 10 m/s over geometrically challenging courses, it achieved a 25.3% reduction in maximum cross-track error, suggesting the spatial representation helps where terrain interactions are most severe.