DeFM: New Depth Foundation Model for Robotics скачать в хорошем качестве

DeFM: New Depth Foundation Model for Robotics

In this AI Research Roundup episode, Alex discusses the paper: 'DeFM: Learning Foundation Representations from Depth for Robotics' DeFM is a self-supervised foundation model trained on a massive dataset of 60 million depth images specifically for robotic applications. By utilizing a self-distillation objective, the model learns geometric and semantic representations that generalize across different sensors and environments. The researchers introduced a unique input normalization strategy to maintain metric awareness, which is crucial for precise physical interactions. The project also offers compact versions of the model, making it suitable for deployment on resource-constrained robotic hardware. It achieves state-of-the-art results in tasks like navigation and manipulation while demonstrating impressive zero-shot transfer from simulation to reality. Paper URL: https://arxiv.org/pdf/2601.18923 #AI #MachineLearning #DeepLearning #Robotics #ComputerVision #DepthSensing #FoundationModels Resources: GitHub: https://github.com/leggedrobotics/defm