MAIA: Anticipating Interaction Anomalies in Contact-Rich Manipulation via Multimodal World Models
MAIA anticipates interaction anomalies in contact-rich manipulation by modeling how robot actions affect future visual and force responses. It uses a Vision–Action–Force World Model to predict these consequences, allowing for proactive anomaly warnings and enabling better action selection without retraining the main policy.