DPA-FTG: Hierarchical Imitation Learning for Contact-Rich Disassembly

A new paper on arXiv (2608.03103) introduces Diffusion Policy Augmented by Fast Trajectory Generation (DPA-FTG), a hierarchical approach to imitation learning for robotic manipulation tasks that require time-varying forces, such as disassembly. The method addresses the trade-off between inference latency and high-frequency control in diffusion policies, which are typically limited in contact-rich tasks like chiseling and prying. DPA-FTG decouples low-frequency planning (5 Hz) from high-frequency force regulation, using a conditional diffusion model to predict latent parameters for trajectory generation. This allows for dynamic interactions while remaining responsive to rapid force transients caused by fracture events. The paper is a cross-type announcement, indicating it may have been presented at a conference and posted on arXiv. The research is relevant to the fields of robotics and artificial intelligence, particularly in industrial automation and maintenance.