Reproduce “Metric-Free Exploration for Topological Mapping by Task and Motion Imitation in Feature Space” on OpenResearch

About this paper

Yuhang He, Irving Fang, Yiming Li, Rushi Bhavesh Shah, Chen Feng

DeepExplorer introduces a metric-free approach for robotic exploration and topological map construction by performing task and motion planning entirely within image feature space. The method, trained with deeply-supervised imitation learning, achieves robust zero-shot generalization in both simulated and real-world environments while demonstrating high data efficiency.

Key insights

• Planning and acting entirely within image feature space, free from metric information, allows for a lightweight and robust exploration system with strong generalization capabilities.
• Deeply-supervised learning, by providing ground-truth supervision at intermediate layers for both feature hallucination and action prediction, significantly enhances the training efficiency and effectiveness of the network.
• The TAMP approach, which decouples high-level goal feature hallucination from low-level action execution, enables efficient and purposeful exploration without explicit spatial coordinates.

Summary by alphaXiv.