Lifelong Embodied Navigation Learning

ICLR 2026 Conference SubmissionAnonymous Authors
Embodied NavigationLifelong LearningRobotics Learning
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 knowledge, we propose an expert subspace orthogonality constraint together and a navigation-specific chain-of-thought reasoning mechanism to capture specific knowledge and enhance instruction-style understanding. Extensive experiments demonstrate the superiority of Uni-Walker for building universal embodied navigation agents with lifelong learning. We also provide the code of this work in the Supplementary Materials.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

The paper introduces Uni-Walker, a framework for lifelong embodied navigation learning (LENL) that addresses catastrophic forgetting when agents adapt to sequential navigation tasks across multiple scenes and instruction styles. It resides in the 'Sequential Environment Adaptation in VLN' leaf, which contains only three papers total, including this work and two siblings. This represents a relatively sparse research direction within the broader taxonomy of 50 papers across 36 topics, suggesting the specific problem of continual task adaptation in vision-language navigation remains underexplored compared to adjacent areas like object navigation or SLAM-based approaches.

The taxonomy reveals that neighboring branches address related but distinct challenges. The sibling leaf 'Cross-Modal Grounding and Reinforcement Learning for VLN' focuses on RL-based vision-language alignment without explicit continual learning mechanisms, while 'User Feedback and Test-Time Adaptation in VLN' emphasizes environment-specific refinement rather than sequential task acquisition. Broader branches like 'Lifelong Learning Architectures for Task and Motion Planning' and 'Continual Reinforcement Learning for Adaptive Navigation' tackle lifelong learning but typically outside the vision-language navigation context. The paper's positioning suggests it bridges VLN instruction-following with continual learning mechanisms, occupying a niche between language-grounded navigation and general lifelong learning frameworks.

Among 30 candidates examined across three contributions, none were identified as clearly refuting the work. The LENL problem formulation examined 10 candidates with no refutations, suggesting the specific framing of sequential VLN task adaptation may be novel within the limited search scope. The Uni-Walker framework with DE-LoRA and the knowledge inheritance/co-activation strategies each examined 10 candidates with similar results. However, this reflects the bounded nature of the semantic search rather than exhaustive coverage; the analysis does not rule out relevant prior work in broader continual learning or parameter-efficient fine-tuning literature outside the top-30 matches.

Given the limited search scope and sparse taxonomy leaf, the work appears to occupy a relatively unexplored intersection of vision-language navigation and continual learning. The absence of refutations among examined candidates suggests potential novelty in the specific problem formulation and architectural choices, though the analysis cannot confirm whether similar mechanisms exist in adjacent research areas not captured by the top-30 semantic matches. The sparse leaf population and lack of overlapping prior work within the examined set indicate this direction warrants further investigation.

Taxonomy

Core-task Taxonomy Papers
50
3
Claimed Contributions
30
Contribution Candidate Papers Compared
0
Refutable Paper

Research Landscape Overview

Core task: lifelong embodied navigation learning with continual task adaptation. The field addresses how autonomous agents can navigate and adapt across evolving environments and tasks without catastrophic forgetting. The taxonomy reveals a rich structure spanning ten major branches. Some branches focus on specific modalities and problem settings: Continual Learning Frameworks for Vision-Language Navigation (e.g., Vision Language Continual[3], Continual VLN[4]) tackle instruction-following in changing environments, while Continual Object Navigation and Goal-Directed Tasks emphasize target-driven exploration. Other branches address architectural and algorithmic foundations, such as Lifelong Learning Architectures for Task and Motion Planning (e.g., Lifelong TAMP[10]) and Continual Reinforcement Learning for Adaptive Navigation (e.g., Continual RL Navigation[38]). Federated and Distributed Lifelong Learning for Robot Navigation (e.g., Lifelong Federated RL[7]) explores multi-agent and decentralized settings, whereas Lifelong Topological and Spatial Mapping for Navigation (e.g., Lifelong Topological Visual[18], Continual SLAM[21]) focuses on persistent spatial representations. Cross-cutting mechanisms and specialized applications round out the landscape. Several active lines of work reveal key trade-offs between task-specific adaptation and general-purpose learning. Vision-language navigation methods like Vision Language Continual[3] and Continual VLN[4] emphasize sequential environment adaptation, balancing linguistic grounding with spatial memory. In contrast, imitation-based approaches such as Continual Imitation Benchmark[1] and Continual Embodied Agents[2] prioritize learning from demonstrations across diverse tasks. Lifelong Embodied Navigation[0] sits within the vision-language continual learning branch, specifically addressing sequential environment adaptation. Compared to Vision Language Continual[3] and Continual VLN[4], which also tackle instruction-following in evolving settings, Lifelong Embodied Navigation[0] appears to integrate task adaptation more explicitly, bridging the gap between environment-specific tuning and broader continual learning mechanisms. This positioning highlights ongoing questions about how to retain knowledge across tasks while remaining responsive to new environmental demands.

Claimed Contributions

Lifelong Embodied Navigation Learning (LENL) problem and benchmark

The authors formalize a new problem setting where embodied navigation agents must continually adapt to sequential navigation tasks spanning multiple scenes and diverse user instruction styles (VLN, OLN, DUN) while retaining previously learned knowledge. They construct a benchmark with 18 navigation tasks for evaluation.

10 retrieved papers
Uni-Walker framework with Decoder Extension LoRA (DE-LoRA)

The authors introduce Uni-Walker, a lifelong embodied navigation framework that uses a novel Decoder Extension LoRA architecture to explicitly separate navigation knowledge into shared components (learned via subspace A) and task-specific components (learned via expert subspaces B), enabling continual learning without catastrophic forgetting.

10 retrieved papers
Knowledge Inheritance Strategy (KIS) and Experts Co-Activation Strategy (ECAS)

The authors design two strategies for learning shared navigation knowledge: KIS initializes new expert subspaces using principal component analysis of previously learned experts with the same instruction style, while ECAS activates multiple related experts during training to exploit shared knowledge and enable smooth knowledge consolidation.

10 retrieved papers

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

Lifelong Embodied Navigation Learning (LENL) problem and benchmark

The authors formalize a new problem setting where embodied navigation agents must continually adapt to sequential navigation tasks spanning multiple scenes and diverse user instruction styles (VLN, OLN, DUN) while retaining previously learned knowledge. They construct a benchmark with 18 navigation tasks for evaluation.

Contribution

Uni-Walker framework with Decoder Extension LoRA (DE-LoRA)

The authors introduce Uni-Walker, a lifelong embodied navigation framework that uses a novel Decoder Extension LoRA architecture to explicitly separate navigation knowledge into shared components (learned via subspace A) and task-specific components (learned via expert subspaces B), enabling continual learning without catastrophic forgetting.

Contribution

Knowledge Inheritance Strategy (KIS) and Experts Co-Activation Strategy (ECAS)

The authors design two strategies for learning shared navigation knowledge: KIS initializes new expert subspaces using principal component analysis of previously learned experts with the same instruction style, while ECAS activates multiple related experts during training to exploit shared knowledge and enable smooth knowledge consolidation.

Lifelong Embodied Navigation Learning | Novelty Validation