Can robots learn to fold clothes from mixed demonstrations?

Household robots struggle with deformable objects like clothing because existing systems train separate policies for each item type. DeMaVLA combines a vision-language backbone with an efficient action expert (using pruned transformers and flow matching) and trains on 5,000 hours of real dual-arm demonstrations plus corrective trajectories from failed attempts. The result: a single policy that folds different clothing items across varying materials and scenes, validated on both simulation and real robot experiments.