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

AI summary

Overview

Research area: Embodied AI, specifically continual learning for object-goal navigation in open-world 3D environments, evaluated on the Habitat simulator with HM3D and MP3D scenes.

Technical level: Advanced (assumes familiarity with imitation learning, knowledge distillation, replay-based continual learning, and multimodal navigation architectures).

Scope: The paper introduces a benchmark for continual object-goal navigation and proposes C-Nav, a framework that lets navigation agents learn new object categories sequentially without forgetting old ones, while storing far less data than naive trajectory replay.

What This Paper Is About

Navigation agents are usually trained once on a fixed list of object categories, but a robot deployed in a real home or office must keep learning new objects (a kettle, a wheelchair, a fire extinguisher) without losing the ability to find things it already knew. Naively fine-tuning on new categories causes catastrophic forgetting: the model drifts in how it represents the world and degrades in how it chooses actions. This paper builds a formal benchmark for that problem and a method that solves it without hoarding full training trajectories.

Key Contributions

1. Continual-ObjectNav benchmark. A new evaluation protocol built on HM3D and MP3D with the Habitat simulator, where object categories are split into four disjoint stages and the agent is tested after each stage on the cumulative set of categories. It supports four mainstream architectures (RNN, BEV, Transformer, LLM-based) and several standard continual-learning baselines (fine-tuning, LoRA, LwF, model merging, data replay).
2. C-Nav, a dual-path anti-forgetting framework. Combines feature-level knowledge distillation across old and new multimodal encoders (representation consistency) with feature replay inside the action decoder (policy consistency), avoiding the storage and privacy costs of retaining raw trajectories.
3. Adaptive experience selection via outlier detection. Treats keyframe selection as an outlier-detection problem in CLIP feature space and uses the Local Outlier Factor (LOF) algorithm to keep only semantically salient frames, such as goal discovery and spatial transitions.
4. Broad empirical validation. Demonstrates consistent gains over baseline continual-learning methods across four model architectures and two datasets, sometimes outperforming baselines that retain every trajectory at greatly reduced memory cost.

Main Findings

• Fine-tuning collapses. Naive sequential fine-tuning drops roughly 40% in success rate on previously learned categories by the final stage across all four architectures, confirming that off-the-shelf navigation policies are highly prone to forgetting.
• C-Nav beats full data replay. Across four architectures, C-Nav improves average success rate by 3.35% on MP3D and 2.75% on HM3D relative to data replay, while storing far less information (compressed features rather than raw trajectories).
• Both anti-forgetting paths are essential. Removing feature distillation costs an average of 22% (HM3D) and 16% (MP3D) success rate; removing feature replay costs 12% and 10%. The two mechanisms are complementary — one stabilizes the encoder, the other stabilizes the decision policy.
• Adaptive selection outperforms uniform sampling. At half-length trajectories, adaptive keyframe selection gains 3.65% SR on HM3D and 3.2% on MP3D over uniform sampling, and remains competitive while using as little as 20% of the memory of uniform or clustering-based sampling.
• Architecture matters but C-Nav transfers. BEV-based decoders consistently outperform the others; LLM-based decoders show no clear advantage, likely due to limited training data, though LoRA helps them disproportionately.
• Robust to stage ordering. Testing three different orderings of the four category stages produces minimal performance variation, which matters for unpredictable real-world deployments.
• Beats zero-shot VLFM in the long run. A frozen vision-language zero-shot system (VLFM) stays stable but cannot improve; C-Nav starts lower in early stages but keeps improving and eventually surpasses it.

Methodology in Plain English

The agent is split into two parts: a multimodal encoder that turns RGB, depth, pose, and the goal object name into a feature vector, and an action decoder (RNN, BEV, Transformer, or LLM) that turns a history of those features into navigation actions. Training happens one stage at a time, with each stage adding new goal categories never seen before.

To stop the encoder from drifting, C-Nav keeps a frozen copy of the previous stage's encoder and penalizes the current encoder whenever its features diverge from the frozen one on the same input — a standard knowledge-distillation trick applied at the feature level rather than the output level.

To stop the decoder from forgetting, C-Nav stores a small set of features (not raw images or trajectories) from earlier stages along with their expert action labels, and replays them during training of the new stage. Because features are compact and privacy-preserving, this sidesteps both the storage blowup and the privacy concerns of storing raw camera streams. Replayed transitions are weighted more heavily when the expert action changes from the previous step, focusing learning on decision points.

The clever part is choosing which frames to keep. Long navigation trajectories are visually redundant — dozens of near-identical hallway frames. C-Nav encodes each frame with CLIP, then runs the Local Outlier Factor algorithm over the cosine-distance neighborhood structure. Frames with LOF scores above 1 are statistically unusual relative to their neighbors; these tend to be the moments when the agent enters a new room, spots the target, or needs to choose a direction. Only those keyframes are stored, encoded, and replayed later.

Why This Matters

Impact on research. Continual learning has been studied extensively for image classification and, more recently, for multimodal LLMs, but the paper argues this is the first formal treatment in object-goal navigation — a domain where forgetting is more damaging because it affects long-horizon behavior, not just a single label. The benchmark gives the community a shared yardstick, and the dual-path design offers a reusable template for other embodied tasks with multimodal inputs.

Real-world applications:

• Home service robots that must learn to find newly introduced household objects (a user's specific mug, a new pet accessory) without forgetting where the kitchen or bathroom is.
• Warehouse and logistics robots whose target inventory changes every time a new product line is added.
• Assistive robots in hospitals or care facilities where new instrument categories appear over time and retraining from scratch is impractical.
• Any long-lived deployed agent that cannot store or transmit raw camera footage for privacy or bandwidth reasons — the feature-only replay design is directly relevant here.

Industry relevance. Deploying continual navigation without retaining raw trajectories reduces both cloud storage costs and privacy/regulatory exposure, which are real barriers for consumer robotics. The finding that gains hold across RNN, BEV, Transformer, and LLM backbones means the technique is not tied to a single vendor's architecture, making it easier to adopt incrementally.

Future Directions

• From simulation to physical robots. The authors explicitly note that dynamic lighting, sensor noise, and hardware variability are untested; validating on real platforms is the natural next step.
• Breaking the linear memory growth. Storage still scales with the number of tasks. Generative replay or trajectory-free approaches that synthesize pseudo-experiences could eliminate the buffer entirely.
• Better handling of very long-tail categories. Stages 2-4 in HM3D contain only one category each; scaling to many small, rapidly arriving categories is an open question.
• Integrating with test-time adaptation and self-evolving world models (e.g., the FSTTA and NavMorph lines of work) rather than treating continual training and test-time adaptation as separate problems.

Target Audience

Robotics and embodied AI researchers working on navigation, continual learning specialists interested in extending their methods beyond classification to sequential decision-making, and applied engineers at robotics companies who need long-lived deployed agents to keep learning without retraining from scratch or retaining sensitive raw sensor data. Readers without a continual-learning or imitation-learning background will need to consult the cited background work to follow the technical sections.

Authors’ 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.