Hierarchical Image-Goal Navigation in Real Crowded Scenarios

Qiaoyun Wu, Jun Wang, Jing Liang, Xiaoxi Gong, Dinesh ManochaarXiv 2021 · OpenReview

Abstract

Image-goal navigation requires an agent to reach a target location defined by a single image within an unfamiliar environment. Existing end-to-end reinforcement learning (RL) approaches often suffer from poor generalization and inefficient exploration in complex, real-world settings. We present a modular hierarchical framework that decomposes the navigation task into mapping, long-term goal prediction, and reactive motion planning. This decoupling allows for the integration of robust real-world components like Gmapping and CrowdMove. Experimental results demonstrate that our approach significantly outperforms state-of-the-art methods, achieving a 17% increase in success rate and a 23% reduction in collisions in real-world crowded scenarios.

Introduction

In the task of Image-Goal Navigation, a robot must navigate to a location represented by a goal image without prior knowledge of the environment's map or its own relative coordinates. This differs from point-goal navigation, which provides precise geometric coordinates. Success in this domain requires the agent to recognize visual landmarks, explore efficiently, and avoid obstacles.

Current state-of-the-art models often employ end-to-end RL, where a single neural network maps raw pixels directly to motor commands. However, these "black box" models frequently overfit to simulation environments and struggle with the computational complexity of simultaneous mapping, planning, and obstacle avoidance, leading to jerky movements and failure in cluttered real-world spaces.

Methods

Our proposed framework addresses these challenges through a modular architecture that separates high-level strategy from low-level control. The system consists of four specialized modules:

  1. Automatic Mapping Module: Generates an 8-channel obstacle map. It utilizes Neural SLAM in simulation and Gmapping in real-world deployments to ensure robust spatial awareness.
  2. Long-term Goal Prediction Module: Acts as the high-level planner. An RL-based policy processes the current map and goal image to select a strategic waypoint (long-term goal).
  3. Reactive Motion Planning Module: Handles local navigation to the waypoint. It employs the Fast Marching Method (FMM) in simulation and the CrowdMove algorithm in the real world to navigate safely around dynamic obstacles.
  4. Navigation Ending Prediction Module: A Siamese network that compares the current visual observation with the goal image to determine when the target has been reached.
The modular architecture for image-goal navigation in both simulation (Habitat) and real-world environments.

By decoupling these tasks, the robot can maintain a consistent exploration strategy while utilizing high-frequency controllers for immediate safety and obstacle avoidance.

Results

The framework was evaluated in the Habitat simulator using the Gibson dataset and on a physical TurtleBot2 platform in various cluttered indoor layouts.

Simulation Performance

In the Habitat simulator, our modular approach achieved a Success Rate (SR) of 0.57, surpassing Neural Topological SLAM (0.55) and significantly outperforming end-to-end models like TD-VNM (0.26).

Real-World Performance

The advantages of the modular design were most evident in physical environments. Our method achieved a 17% higher success rate and a 23% lower collision rate compared to previous state-of-the-art models. Qualitatively, our system demonstrated more continuous movement and efficient exploration, whereas baseline models often became stuck or exhibited oscillatory behavior.

Qualitative comparison of navigation trajectories and observations between our method and various baselines.

Discussion and Conclusion

The success of this hierarchical approach highlights the importance of decoupling high-level reasoning from low-level execution. By separating "where to go" from "how to move," the system becomes more resilient to sensor noise and environmental complexity.

Future work will focus on addressing actuation lag in extremely crowded environments and incorporating semantic mapping to leverage object-room associations. Additionally, moving from 2D occupancy maps to 3D perception could further enhance navigation in multi-level or highly complex spaces.

Summary updated

BibTeX
@misc{wu2021hierarchical,
  author = {Qiaoyun Wu and Jun Wang and Jing Liang and Xiaoxi Gong and Dinesh Manocha},
  title = {{Hierarchical Image-Goal Navigation in Real Crowded Scenarios}},
  year = {2021},
  url = {https://openreview.net/forum?id=mhGRTD36FyH},
}
Figure