Trinity: New Model Unifies Terrain Segmentation for Outdoor Robots

Autonomous robots operating in unstructured outdoor environments struggle with terrain understanding because existing methods either require robot-specific annotations or are limited to predefined semantic classes. A team of researchers from multiple institutions—including Müller, Boerdijk, Durner, and others—proposes Trinity, a transformer-based architecture that unifies class-specific semantic segmentation and class-agnostic terrain segmentation within a single network. The key insight is to segment terrain regions purely based on visual appearance, without relying on predefined semantic labels or robot-dependent traversability scores. This allows the model to learn robot-agnostic visual terrain priors that can be combined with platform-specific experience for downstream tasks like traversability estimation, visual odometry, and mission planning. To train such a model at scale, the authors extended the OAISYS simulator to generate RUGDSynth, a synthetic dataset inspired by RUGD that provides class-agnostic terrain samples with diverse appearances. They also collected EXTerra, a real-world dataset with both class-specific and class-agnostic terrain annotations. Experiments demonstrate that the joint segmentation approach outperforms separate pipelines and generalizes effectively to complex outdoor environments. Code and datasets will be released upon publication.