SAFER-CoT: Safety-Aware Robot Foundation Models with Embodied Reasoning via Chain-of-Thought

The first safety-aware robot foundation model that integrates semantic and physical safety reasoning directly into embodied chain-of-thought (CoT) planning. The project combines large-scale generation of safe and deliberately unsafe robot demonstrations with frontier vision-language models to produce fine-grained, temporally consistent safety explanations over video sequences. These annotations are used to train VLA policies that can anticipate, explain, and mitigate unsafe actions before they are executed. Leveraging EuroHPC resources enables scalable safety labelling and training of large VLA models. SAFER-CoT will release open-source datasets, models, and pipelines, advancing explainable and reliable robot learning and supporting safe human-robot interaction in line with EU AI safety goals.