Abstract
We present a novel approach for image-goal navigation, where an agent navigates with a goal image rather than accurate target information, which is more challenging. Our goal is to decouple the learning of navigation goal planning, collision avoidance, and navigation ending prediction, which enables more concentrated learning of each part. This is realized by four different modules. The first module maintains an obstacle map during robot navigation. The second predicts a long-term goal on the real-time map periodically, which can thus convert an image-goal navigation task to several point-goal navigation tasks. To achieve these point-goal navigation tasks, the third module plans collision-free command sets for navigating to these long-term goals. The final module stops the robot properly near the goal image. The four modules are designed or maintained separately, which helps cut down the search time during navigation and improves the generalization to previously unseen real scenes. We evaluate the method in both a simulator and in the real world with a mobile robot. The results in real complex environments show that our method attains at least a $17\%$ increase in navigation success rate and a $23\%$ decrease in navigation collision rate over some state-of-the-art models.
Introduction
In the field of robotics, Point-Goal Navigation is a well-established task where a robot is given specific coordinates to reach. In contrast, Image-Goal Navigation is significantly more complex: the robot is shown a picture of a destination and must find it in an unfamiliar environment without a pre-existing map or global positioning.
Traditional approaches often rely on "end-to-end" deep reinforcement learning (RL). While these models perform well in controlled simulations, they frequently fail in real-world applications due to overfitting. They often struggle with long-range navigation, visual noise, and a lack of robust geometric understanding, leading to frequent collisions. This research addresses these challenges by proposing a modular, hierarchical framework that decouples global planning from local control.
Modular Architecture
The proposed system breaks the navigation task into four specialized, interoperable components:
- Automatic Mapping: The robot builds an explicit 2D obstacle map in real-time. In simulation, this is achieved via Neural SLAM; in real-world deployments, a laser-based Gmapping approach is used.
- Long-term Goal Prediction: An RL-based policy analyzes the current map and the goal image to predict a "waypoint" likely to lead toward the target.
- Reactive Motion Planning: This module translates waypoints into collision-free steering commands. The system uses the Fast Marching Method in simulation and the more robust CrowdMove algorithm for real-world scenarios.
- Navigation Ending Prediction: A Siamese neural network compares the robot's current visual observation with the goal image, issuing a stop command when the robot is within approximately 1 meter of the target.
By separating these tasks, the robot avoids the burden of learning visual recognition, spatial planning, and obstacle avoidance simultaneously, leading to better generalization.
Experimental Results
The model was trained using the Habitat simulator on the Gibson dataset and evaluated in real-world environments using a TurtleBot2 equipped with RPLIDAR and monocular cameras.
Quantitative Performance
In simulation, the proposed method achieved a success rate (SR) of 57%, significantly outperforming the state-of-the-art (SOTA) end-to-end model (TD-RL-ITR) at 44%. In real-world tests, the modular system maintained a success rate between 32% and 58%, whereas end-to-end models often plummeted to success rates as low as 6% in complex, cluttered rooms. Furthermore, the system reduced the collision rate by 23% compared to the best-performing baseline.
Qualitative Analysis
Qualitative comparisons demonstrate that the modular approach produces more efficient and direct paths toward the goal. While baseline models often get stuck in local minima or fail to recognize the target location, the proposed system consistently updates its long-term goals based on the evolving map.
Conclusion
This work demonstrates that decoupling mapping, planning, and control is essential for robust image-goal navigation. By using modular learning, the system achieves superior "sim-to-real" transfer and safety. Future work will focus on incorporating semantic properties—such as identifying doors versus walls—to further enhance high-level planning and mitigate "freezing" behaviors in extremely crowded environments.
Summary updated