This project addresses continual learning in Vision-Language-Action (VLA) models to reduce catastrophic forgetting. Instead of using replay buffers or co-training, it applies constrained optimization to limit parameter drift away from weights optimal for prior tasks.