ORDformer: A Transformer-based Benchmark for Occupancy Prediction in Off-Road Terrain

Autonomous navigation in off-road environments is challenging due to complex terrains and undefined traversability boundaries. This paper introduces ORDformer, a deformable transformer-based network for 3D semantic occupancy prediction that integrates LiDAR geometric data with camera-derived semantic information through deformable cross-attention mechanisms. ORDformer addresses key limitations in existing methods by delivering precise semantic predictions for off-road conditions. Experimental results show that ORDformer outperforms existing models, achieving 50.20% IoU for scene completion and 10.50% mIoU for semantic scene completion. The method successfully integrates multimodal data, improving traversability estimation in diverse environments. Future work includes further integration with 3D planning algorithms for comprehensive trajectory planning and the collection of an expanded off-road dataset containing more complex obstacles to improve model performance.