C-NAV: Towards Self-Evolving Continual Object Navigation in Open World

TL;DR

This work tackles continual object navigation in open-world environments by introducing a continual-ObjectNav benchmark and the C-Nav framework. C-Nav addresses catastrophic forgetting with a dual-path anti-forgetting mechanism that enforces representation consistency via feature distillation and policy stability via feature replay, paired with an adaptive experience sampling strategy based on LOF-driven keyframe selection. Across multiple backbone architectures and two challenging datasets, C-Nav achieves superior performance and substantially lower memory overhead compared to baselines, demonstrating robust continual learning in long-horizon, multimodal navigation. The approach offers practical implications for deploying adaptable embodied agents while mitigating privacy and storage concerns inherent in naive trajectory replay.

Abstract

Embodied agents are expected to perform object navigation in dynamic, open-world environments. However, existing approaches typically rely on static trajectories and a fixed set of object categories during training, overlooking the real-world requirement for continual adaptation to evolving scenarios. To facilitate related studies, we introduce the continual object navigation benchmark, which requires agents to acquire navigation skills for new object categories while avoiding catastrophic forgetting of previously learned knowledge. To tackle this challenge, we propose C-Nav, a continual visual navigation framework that integrates two key innovations: (1) A dual-path anti-forgetting mechanism, which comprises feature distillation that aligns multi-modal inputs into a consistent representation space to ensure representation consistency, and feature replay that retains temporal features within the action decoder to ensure policy consistency. (2) An adaptive sampling strategy that selects diverse and informative experiences, thereby reducing redundancy and minimizing memory overhead. Extensive experiments across multiple model architectures demonstrate that C-Nav consistently outperforms existing approaches, achieving superior performance even compared to baselines with full trajectory retention, while significantly lowering memory requirements. The code will be publicly available at https://bigtree765.github.io/C-Nav-project.

Figures (15)

• Figure 1: Continual Object Navigation: The robot must continually learn from new data while retaining its ability to navigate to previously seen object goals. After training across multiple phases, the baseline model suffers from representation drift and policy degradation, leading to a loss of navigation capability on previously learned tasks, whereas C-Nav enables cumulative knowledge retention across phases.
• Figure 2: Different navigation model architectures. According to the action decoder, they can be categorized into a) RNN-Based, b) Bev-Based, c) Transformer-Based, and d) LLM-Based.
• Figure 3: Overview of the proposed C-Nav framework for continual object navigation. It consists of two key components: (1) adaptive experience selection, which identifies semantically meaningful keyframes via LOF in the representation space, reducing storage and redundancy; and (2) dual-path anti-forgetting, which mitigates representation and policy drift through a feature distillation path (bottom left) and a feature replay path (bottom right), ensuring stability across encoder and decoder modules. For each task, C-Nav retains $p$ sparse trajectory features in the deep feature space.
• Figure 4: Results of SR and SPL on the MP3D and HM3D datasets. Solid lines represent SR, while dashed lines represent SPL.
• Figure 5: Ablation on experience selection and retention ratio on MP3D with a Transformer-based navigation architecture.