Active Perception Meets Rule-Guided RL: A Two-Phase Approach for Precise Object Navigation in Complex Environments

Video Demos

Video 1: Comparative Analysis with Baseline Methods

demo1

Video 2: Active Navigation in Last-Mile Phase

demo2

Environment Setup

This repository presents a streamlined implementation of the original Active Perception with Rule-Guided Reinforcement Learning (APRR) approach. Redundant components from the original formulation have been omitted to enhance deployability and interpretability, while maintaining performance.

1.Follow the environment configuration guidelines for HomeRobot as specified in the official documentation: HomeRobot Setup Instructions

2.Alternatively, utilize the Docker image provided by the HomeRobot project: HomeRobot Docker Repository

3.Ensure the Conda environment named "home-robot" is properly initialized:

conda activate home-robot
cd home-robot/src/home-robot
pip install -e .

4.Install YOLO-World and SAM (Segment Anything Model) following the guidelines in the Ultralytics documentation: Ultralytics Installation Guide

5.Configure simulation

cd home-robot
bash set.sh

Evaluation

The pre-trained model checkpoint is available for download at: Model Checkpoint (BaiduNetDisk)

To evaluate the baseline method:

bash runb.sh

To evaluate the APRR framework:

bash run_rl.sh

Training Procedures

For detailed training protocols, please refer to the official HomeRobot documentation: HomeRobot Training Guidelines