Imitation learning with action chunking has achieved strong results in robotic manipulation by predicting and executing action sequences from vision and pose observations. However, in contact-rich tasks, policies trained only under vision- and positio...
Imitation learning with action chunking has achieved strong results in robotic manipulation by predicting and executing action sequences from vision and pose observations. However, in contact-rich tasks, policies trained only under vision- and position-based supervision struggle to regulate contact forces, often leading to unstable or unsafe interactions. This motivates incorporating force information so that policies can reason not only about where to move, but also how hard to push. Yet even when force is included, existing chunk-based policies still face critical challenges: they typically operate in an open-loop manner without real-time correction and suffer spatio-temporal discontinuities when successive chunks belong to inconsistent modes, resulting in jerky transitions and abrupt force changes. To address these issues, a Force-Aware Hierarchical Diffusion Policy is introduced, comprising a high-level Global Policy and a low-level Local Policy. The Global Policy, implemented as a high-level diffusion model, predicts long-horizon and force-aware action chunks by explicitly incorporating force information. Meanwhile, the Local Policy serves as a low-level reactive controller that predicts fine-grained actions in real time, mitigating spatio-temporal discontinuities at chunk boundaries and enabling rapid adaptation to environmental changes and unexpected disturbances during execution. In real-world contact-rich tasks, our approach outperforms baseline chunk-based visuomotor policies in terms of stable contact, transition smoothness, and responsiveness.