Reproduce “Instance-aware Exploration-Verification-Exploitation for Instance ImageGoal Navigation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xiaohan Lei, Min Wang, Wengang Zhou, Li Li, Houqiang Li
The research from the University of Science and Technology of China introduces the Instance-aware Exploration-Verification-Exploitation (IEVE) framework, which significantly improves Instance ImageGoal Navigation by incorporating a human-inspired verification stage. This modular framework achieved 0.702 Success and 0.252 SPL on the HM3D-SEM dataset, establishing a new state-of-the-art and surpassing previous methods by a margin of 0.141 in Success.
Key insights
- Incorporating a 'Verification' stage, inspired by human decision-making, is crucial for improving instance-level identification accuracy, allowing agents to confirm potential targets before committing.
- Adaptive thresholds for local feature matching, which adjust based on the agent's distance to a potential target, are essential for reliable instance re-identification in varying observation conditions.
- A modular framework design, which integrates state-of-the-art perception models (e.g., instance segmentation) and specific navigation policies, provides robustness and computational efficiency over end-to-end learning for complex embodied tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.