RECON

将潜在目标模型与拓扑图相结合，RECON 可以在新环境中快速发现用户指定的目标并可靠地导航到它们。我们的方法包括三个阶段：
先前经验：目标条件距离和动作模型是使用先前访问过的环境中的经验进行训练的。通过使用时间步长作为距离的代理和事后重新标记方案来获得对模型训练的监督。
探索新环境：当进入新环境时，RECON 结合使用基于前沿的探索和潜在目标采样与学习模型来发现视觉目标。学习到的模型也针对新环境进行了微调。
导航探索的环境：给定一个探索的环境（由拓扑图G表示）和模型，RECON 使用G通过规划图中的子目标路径来导航到目标图像。

Combining latent goal models with topological maps, RECON enables rapid discovery of user-specified targets and reliable navigation to them in novel environments. Our method consists of three stages:
Prior Experience: Goal-conditioned distance and action models are trained using experiences from previously visited environments. Supervision for model training is obtained by using time steps as a proxy for distance and an offline relabeling scheme.
Exploring Novel Environments: Upon entering a new environment, RECON combines frontier-based exploration and latent goal sampling with the learned models to discover visual targets. The learned models are also fine-tuned for the novel environment.
Navigating Explored Environments: Given an explored environment represented by the topological map G and the learned models, RECON uses G to navigate to the target image by planning subgoal paths within the graph.

搜集汇总

背景与挑战

背景概述

RECON是一个专注于机器人学习未知环境的数据集，由卡内基梅隆大学和加州大学伯克利分校于2021年发布。其核心方法结合潜在目标模型与拓扑图，通过三个阶段（先前经验训练、探索新环境和导航探索的环境）实现快速目标发现和可靠导航，适用于视觉目标导向的机器人应用。