Diffusion Trajectory-guided Policy: A Novel Framework for Long-Horizon Robot Manipulation
Keywords: Robotics, Imitation Learning, Generative Model, Vision-Language Action
Abstract: Recently, Vision-Language Models (VLMs) have made substantial progress in robot imitation learning, benefiting from increased amounts of demonstration data. However, the high cost of data collection remains a significant bottleneck, and the scarcity of demonstrations often result in poor generalization of the imitation policy, especially in long-horizon robotic manipulation tasks. To address these challenges, we propose the Diffusion Trajectory-guided Policy (DTP) framework, which generates task-relevant trajectories through a diffusion model to guide policy learning for long-horizon tasks. Furthermore, we demonstrate that our DTP method offers a useful interface for prompt engineering, providing a novel way to connect robot manipulation skills with interactions involving LLMs or humans. Our approach employs a two-stage training process: initially, we train a generative vision-language model to create diffusion task-relevant trajectories, then refine the imitation policy using these trajectories. We validate that the DTP method achieves substantial performance improvements in extensive experiments on the CALVIN simulation benchmark, starting from scratch without any external pretraining. Our approach outperforms state-of-the-art baselines by an average of 25% in success rate across various settings.
Primary Area: applications to robotics, autonomy, planning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 4183
Loading