Event

MAIA anticipates contact-rich manipulation anomalies, raising task success rate to 61.5%

1 reports · 1 sources

Yi Qiao et al. propose MAIA, a multimodal anomaly anticipation framework for contact-intensive manipulation: its vision-action-force world model learns normal interaction dynamics and predicts the visual and force responses of planned actions, while a lightweight anomaly head predicts upcoming interaction anomalies. The authors also release CRIA, a real multimodal dataset containing 2,852 trajectories across 18 contact-intensive manipulation tasks, with frame-by-frame annotations of anomaly onset and evolution. Experimental results reported in the paper show that on seen and new tasks, MAIA's anomaly discrimination outperforms representative baselines, with bACC of 0.765 and 0.710, respectively, and median warning lead times of 28.0 and 25.1 steps; using its predictions for anomaly-aware action selection, without retraining the underlying manipulation policy, raises end-to-end task success rate from 52.6% to 61.5%.