HDFlow: Hierarchical Diffusion-Flow Planning for Long-horizon Tasks

Generative models have shown promise for long-horizon planning in robotics, but existing approaches often lack principled hierarchical decomposition and suffer from high computational costs due to iterative denoising. To address this, a team from Peking University introduces HDFlow (Hierarchical Diffusion-Flow), which optimally combines two generative paradigms: diffusion models for high-level exploration and rectified flow models for low-level trajectory generation. The high-level diffusion planner generates sequences of strategic subgoals in a learned latent space, leveraging diffusion's powerful exploratory capabilities. These subgoals then guide a low-level rectified flow planner that produces smooth, dense trajectories using an ordinary differential equation (ODE)-based process, which is faster and more efficient than iterative denoising.

HDFlow was evaluated on four challenging furniture assembly tasks (both in simulation and real-world), where it significantly outperformed state-of-the-art methods. The team also demonstrated generalizability on two long-horizon benchmarks covering diverse locomotion and manipulation tasks. The work has been accepted as a Spotlight paper at ICML 2026. By effectively merging the strengths of diffusion and flow models, HDFlow offers a principled and efficient framework for complex, long-horizon robot planning that could accelerate deployment in manufacturing, home assistance, and other real-world domains.