DeFM: Learning Foundation Representations from Depth for Robotics

DeFM is a self-supervised foundation model trained entirely on depth images for robotic applications. Using a DINO-style self-distillation objective on a curated dataset of 60M depth images, DeFM learns geometric and semantic representations that generalize to diverse environments, tasks, and sensors. It achieves state-of-the-art performance on depth-based benchmarks and is released for direct use in robotic learning without task-specific fine-tuning.