Lifelong Embodied Navigation Learning
Xudong Wang, Jiahua Dong, Baichen Liu, Qi Lyu, Lianqing Liu, Zhi Han
TL;DR
This paper introduces Uni-Walker, a framework enabling embodied navigation agents to learn multiple tasks continually without forgetting, by decoupling shared and task-specific knowledge and employing novel transfer and reasoning strategies.
Contribution
The paper presents Uni-Walker, a novel lifelong embodied navigation framework that effectively separates shared and task-specific knowledge for continual learning.
Findings
Uni-Walker outperforms existing methods in lifelong navigation tasks.
The knowledge inheritance and experts co-activation strategies improve knowledge transfer.
Navigation-specific reasoning enhances instruction understanding.
Abstract
Embodied navigation agents powered by large language models have shown strong performance on individual tasks but struggle to continually acquire new navigation skills, which suffer from catastrophic forgetting. We formalize this challenge as lifelong embodied navigation learning (LENL), where an agent is required to adapt to a sequence of navigation tasks spanning multiple scenes and diverse user instruction styles, while retaining previously learned knowledge. To tackle this problem, we propose Uni-Walker, a lifelong embodied navigation framework that decouples navigation knowledge into task-shared and task-specific components with Decoder Extension LoRA (DE-LoRA). To learn the shared knowledge, we design a knowledge inheritance strategy and an experts co-activation strategy to facilitate shared knowledge transfer and refinement across multiple navigation tasks. To learn the specific…
Peer Reviews
Decision·ICLR 2026 Poster
- Novel Problem: The paper formally introduces and tackles the problem of lifelong learning for embodied navigation, a crucial challenge for developing general-purpose agents. - Comprehensiv Method: Uni-Walker is a well-thought-out framework that addresses the core challenges of knowledge transfer and catastrophic forgetting by explicitly decoupling and managing shared and task-specific knowledge. - Rigorou Evaluation: The experimental validation is a major strength. The creation of a new benchm
System Complexity: The final Uni-Walker model is a composite of many components (KIS, ECAS, SSC, ESOC, NSCoT, TAKA). While the ablation studies justify each part's inclusion, the overall system is quite complex. A discussion on the relative importance of these components or potential simplifications would be beneficial.
1. The paper introduces a novel problem (LENL) and a corresponding benchmark, which is a vital contribution. The Uni-Walker framework, with its DE-LoRA and specialized strategies (KIS, ECAS, ESOC, NSCoT, TAKA), demonstrates high originality in combining and extending existing ideas for the specific challenges of embodied lifelong learning. 2. The empirical results are robust and convincing. Uni-Walker achieves state-of-the-art performance, significantly reducing catastrophic forgetting compared
1. While each component of Uni-Walker (DE-LoRA, KIS, ECAS, SSC, ESOC, NSCoT, TAKA) is well-justified, the overall framework is quite intricate with multiple interacting mechanisms. It might be challenging to disentangle the precise individual contributions or to simplify the architecture without performance degradation. A discussion on the trade-offs between complexity and performance, or potential avenues for simplification, could be beneficial. 2. While the new benchmark is excellent, 18 tasks
* Novelty In this work, the author propose a kind of new task in navigation, namely, lifelong embodied navigation learning. This task itself is interesting and practical for embodied navigation in the real world. * Clarity The paper is with good structure. Hence, the clarity is basically good. * Significance This paper focuses on a new task of lifelong embodied navigation learning. It would possibly bring new research works for navigation. Hence, the significance is basically OK for this c
* Method Even though the task is kind of new, the techniques to address this task are relatively straightforward. It basically leverages LORA adaption in the decoder for decoupling navigation knowledge into shared and specific parts. Then, they integrate the existing strategies to learn these two types of knowledge for life long learning. Please futher clarify the key novel design in the proposed framework, instead of adapting and integrating the straightforward strategies together on a newly-d
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsMultimodal Machine Learning Applications · Domain Adaptation and Few-Shot Learning · Social Robot Interaction and HRI