In contact-rich manipulation, actions that look equally plausible from vision alone can lead to very different contact outcomes, and reactive control based on current tactile feedback often responds too late to prevent failure. ForeTac, a predictive contact-guidance framework, tackles this by giving generative robot policies a sense of foresight: it predicts what tactile feedback will result from candidate actions before any motion is executed.
The framework operates as a represent-predict-evaluate-guide loop. It encodes the current tactile state, predicts future tactile latent trajectories conditioned on candidate action sequences, and evaluates each predicted trajectory with a contact-quality objective learned from successful and failed physical interactions. At inference, gradients of that objective are backpropagated through the tactile predictor to steer action generation, while all model parameters remain frozen.
ForeTac was evaluated on four real-robot tasks—planar and curved-surface wiping, card swiping, and fragile-object grasping—using both diffusion and flow-matching policies. Compared with a vision-only Diffusion Policy baseline, ForeTac improved the average task success rate from 36.3% to 71.3% and the average contact success rate from 20.0% to 66.3%, and it also produced gains with the flow-matching policy.
Component analyses support the key design choices: the learned tactile representation, action-conditioned foresight, and generation-time gradient guidance all contribute to the gains, and predicted tactile latent trajectories preserve contact-quality ordering. The results suggest that predictive touch can be integrated into existing generative policies without retraining, offering a practical path toward more reliable contact-rich manipulation.