ABSTRACT

Robotic manipulation often requires precise calibration and localization, resulting in long setup times and poor adaptability to new environments. While end-to-end learning from human demonstrations can reduce programming and calibration efforts, collecting large datasets for each task is impractical due to time and labor constraints. To address this challenge, we introduce Goal-Directed Visual Diffusion Policy, which generates synthetic demonstration data using multiple single-shot camera observations of a goal location. By integrating traditional motion planning with diffusion policy, this approach significantly reduces the need for task demonstrations while maintaining comparable accuracy. Validation through both simulations and lab experiments demonstrates its effectiveness for real-world applications. The open-source implementation is available at https://github.com/DinhDangKhoaLe/GoalDirected-Diffusion-Policy.