DexTacWAM models multi-finger contact for dexterous robot manipulation

Berkeley AI presented DexTacWAM, a model that adapts a pretrained video world model to predict multi-finger contact and generate actions. It reports the highest score on each of six contact-rich manipulation tasks. In a four-task comparison using the same tactile encoder, policy architecture and training procedure, replacing the model’s touch prediction with direct tactile features lowered the mean score from 74.7 to 26.6. Berkeley AI says the code, model weights and datasets are open-source.