Event
DeFM: A Robot Foundation Model Trained Only on Depth Images
1 reports · 1 sources
Manthan Patel and co-authors propose DeFM, a robot foundation model trained entirely via self-supervision on depth images, and have released the pretrained weights. The authors state that the model uses a DINO-style self-distillation objective to learn geometric and semantic representations from 60 million depth images, and can generalize across different environments, tasks, and sensors. The authors also state that these pretrained models can be used directly for depth-based robot learning without task-specific fine-tuning; they additionally distilled DeFM into a compact model for use in resource-constrained robotic systems.
Report timeline
- DeFM: A Foundation Model for Learning Robotic Representations from Depth Images
DeFM is a foundation model trained entirely on depth images through self-supervised learning. It uses a DINO-style self-distillation objective to learn geometric and semantic representations from a dataset of 60 million depth images that can generalize across different environments, tasks, and sensors.
arXiv cs.RO ·