ForceLDM: A Force-aware Latent Dynamics Model for Contact-Rich Manipulation

This research presents ForceLDM, a force-aware latent dynamics model designed to improve contact-rich manipulation by learning future interaction dynamics through knowledge distillation. The method employs a teacher network with privileged future force/torque and optical flow signals to extract dynamic features, which are distilled into a student network trained only on current observations. This allows the deployed student model to reason about upcoming contacts without relying on future sensory inputs. Additionally, a curriculum-based progressive noise injection strategy is introduced to enhance robustness and generalization. Experiments on five real-world tasks show that ForceLDM outperforms state-of-the-art baselines, generalizes to novel objects, and remains robust to environmental perturbations. Project page and code are available at https://tpy001.github.io/ForceLDM/.