Robots need to forget dangerous moves without forgetting how to see — researchers build the first method that works

Foundation models that control robots (vision-language-action models) now face a new problem: you can't safely remove a dangerous behavior by editing just the vision part or just the language part, because the dangerous knowledge is spread across both. Researchers built VLA-Forget, a method that surgically removes unsafe behaviors from all three layers at once while keeping the robot's ability to see, understand language, and move intact.

Embodied AI systems — robots that see and act in the real world — are moving from research to deployment. If a robot learned to do something dangerous or privacy-invasive, you need to remove that behavior fast without rebuilding the entire system. Before this, unlearning methods worked on standalone vision or language models, but when behavior emerges from the interaction of perception, language grounding, and action layers fused together, you can't just edit one layer and call it safe. VLA-Forget means that safety teams and robot companies can now actually remove specific unsafe behaviors in deployment without the months of retraining that would otherwise be necessary.