Memory-Driven Multimodal Chain of Thought for Embodied Long-Horizon Task Planning
Keywords: Task Planning; Multimodal Large Language Model; Reasoning
Abstract: Existing methods excel in short-horizon tasks but struggle with complex, long-horizon planning in dynamic environments. To address these limitations, we propose the Memory-Driven Multimodal Chain of Thought (MCoT-Memory), a framework designed to enhance task planning through two key innovations: 1) Evolving Scene Graph-Driven Chain of Thought with CoT Memory Retrieval, which enables the agent to continuously update a scene graph with visual information captured along its trajectory, providing a structured and dynamic representation of the environment that informs real-time decision-making, and uniquely incorporates CoT memory retrieval to allow the agent to leverage past experiences in its reasoning process; 2) Stepwise Confidence-Driven Memory Retention, which employs an expert model to evaluate reasoning across multiple dimensions of accuracy, ensuring that only high-confidence experiences are retained in memory for future retrieval, thus enabling the agent to build on valuable insights and improve performance in long-horizon tasks.
To advance long-horizon task planning, we present ExtendaBench, a comprehensive benchmark encompassing 1,198 tasks across two simulators, VirtualHome and Habitat 2.0. The tasks are categorized into ultra-short, short, median, and long tasks. Extensive experiments demonstrate that prior methods struggle with long-horizon tasks, while MCoT-Memory significantly improves performance, marking it as a promising approach for embodied task planning.
Primary Area: applications to computer vision, audio, language, and other modalities
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: 4242
Loading