SafeVLM

Towards Safe LVM with Automatic Masking and Joint Intention Analysis

**SafeVLM** is a research framework for training safety-aware and uncertainty-aware Vision-Language Models (VLMs) using reinforcement learning. It integrates computer vision safeguards, rating-based RL, and VLM reasoning to improve reliability in embodied AI tasks across benchmarks like MetaWorld, ALFRED, and RoboSuite.