Research map / Continual learning / Online and replay-based
C-NAV: Towards Self-Evolving Continual Object Navigation in Open World
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces a continual object navigation benchmark (Continual-ObjectNav on HM3D and MP3D) and C-Nav, a dual-path anti-forgetting framework with feature distillation, feature replay, and LOF-based adaptive experience selection.
Problem
Object navigation agents trained on static trajectories and fixed object categories suffer catastrophic forgetting when they must learn new object categories continually in open-world settings.
Approach
Combines feature-level distillation of the multimodal encoder (representation consistency) with replay of stored trajectory features and actions in the action decoder (policy consistency), and selects replay frames as LOF outliers in CLIP feature space to cut memory use.
Key results
- On HM3D with the Bev-based decoder, C-Nav reaches 56.3 avg SR and 24.0 avg SPL versus 53.2 SR for data replay.
- On MP3D with the Transformer decoder, C-Nav reaches 38.1 avg SR and 12.1 SPL, outperforming data replay (37.7 SR) while storing fewer frames.
- Removing feature distillation drops HM3D average SR by about 22 points on average across architectures, the largest ablation effect.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use C-Nav instead of data replay | Continual object navigation · HM3D | average SR: +2.75% | Helped |
| use C-Nav instead of data replay | Continual object navigation · MP3D | average SR: +3.35% | Helped |
| add feature distillation (KD) (ablation) | C-Nav · HM3D and MP3D | average SR: 22% and 16% | Helped |
| add feature replay (FP) (ablation) | C-Nav · HM3D and MP3D | average SR: 12% and 10% | Helped |
| replace adaptive sampling with uniform sampling (ablation) | C-Nav · HM3D and MP3D | SR: 3.65% on HM3D; 3.2% on MP3D | Hurt |
| use adaptive sampling instead of full-length feature replay (ablation) | C-Nav · HM3D and MP3D | SR: 1.9% and 1.3% | Hurt |
| use adaptive sampling instead of full data replay (ablation) | C-Nav · HM3D and MP3D | average SR: +0.8% on HM3D; +2.0% on MP3D | Helped |
| fully freeze the multimodal encoder instead of selectively fine-tuning projectors with consistency regularization (ablation) | C-Nav · MP3D Continual ObjectNav | performance | Hurt |
| increase loss weighting coefficient from 5 to 10 (ablation) | C-Nav · MP3D Stage 4 | SR: 36.81 -> 25.29 | Hurt |
| use C-Nav instead of data replay | Continual object navigation · HM3D old tasks Stage 4 | SR: 32.94 -> 42.61 | Helped |
| replace C-Nav with naive fine-tuning | Continual object navigation · HM3D old tasks Stage 2 | SR: 61.38 -> 7.72 | Hurt |
Surprising findings
- Adaptive keyframe selection at 50% length nearly matches full-length feature replay (drop of 1.9 SR on HM3D), suggesting heavy redundancy in navigation trajectories.
- Fully freezing the multimodal encoder degrades performance, which the authors take as evidence that some encoder adaptation is needed.
Limitations
- No validation on physical robots; real-world lighting and sensor noise untested.
- Storage for replay still grows linearly with the number of tasks.
- Only four stages with disjoint categories were evaluated.
Experimental setup
Datasets: HM3D MP3D Habitat-Web demonstrations PIRL demonstrations
Benchmarks: Continual-ObjectNav (HM3D) Continual-ObjectNav (MP3D)
Metrics: Success Rate (SR) SPL
Baselines: Finetuning LoRA LwF Model merge Data replay VLFM
Builds on: LwF LoRA Local Outlier Factor CLIP Habitat
Closest papers by meaning
- Hydra-Nav: Object Navigation via Adaptive Dual-Process Reasoning (ICML 2026 · similarity 0.88)
- NavGPT: Explicit Reasoning in Vision-and-Language Navigation with Large Language Models (AAAI 2024 · similarity 0.87)
- Reasoning in Visual Navigation of End-to-end Trained Agents: A Dynamical Systems Approach (CVPR 2025 · similarity 0.87)
- MemoNav: Working Memory Model for Visual Navigation (CVPR 2024 · similarity 0.87)
- Trajectory Diffusion for ObjectGoal Navigation (NeurIPS 2024 · similarity 0.87)
- An Interactive Navigation Method with Effect-oriented Affordance (CVPR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.