Topometric Imitation Learning For Route Following Under Appearance Change

Shaojun Cai, Yingjia WanWorkshops @ CVPR 2020 · DOI

Introduction

Traditional autonomous navigation systems heavily rely on Simultaneous Localization and Mapping (SLAM) to construct high-definition metric maps. While effective in static environments, SLAM-based approaches often suffer from three primary limitations: fragility under varying lighting or weather conditions, significant storage requirements for dense geometric data, and high computational costs that scale poorly with environment size.

Inspired by biological navigation in mammals, which utilizes landmarks and topological relationships rather than precise geometric coordinates, this paper proposes a two-level navigation architecture. This system combines a topological-metric memory with conditional imitation learning to enable robust route following in complex urban environments, even when the visual appearance changes drastically between the initial "teaching" phase and subsequent "repeating" phases.

Methodology

The proposed framework utilizes a "Topometric" approach that decouples global localization from local control. The architecture is composed of three primary modules: a topological-metric memory, a visual place recognition system, and a local controller.

Topological-Metric Memory

During a single "teach" pass, the system constructs a graph where vertices represent specific locations. Each node stores a visual signature and a navigational command. The authors explore two types of commands:

  1. Discrete Local Move (DLM): High-level instructions such as "turn left" or "go straight."
  2. Local Path (LP): A 2D coordinate path that provides the geometric flexibility needed for complex maneuvers like navigating roundabouts.

Visual Place Recognition and Control

To determine the vehicle's current position on the map, a Siamese Network (based on ResNet50) compares the live camera feed to stored images. By projecting images into a common feature space, the network learns to recognize global scene features that remain consistent despite weather or lighting changes.

The system architecture includes a Siamese network for similarity comparison, a localization module, and an LP controller for steering.

The local controller is a neural network trained via Conditional Imitation Learning (CIL). It maps the current camera image and the retrieved navigational command directly to control outputs (steering and acceleration). To ensure robustness, noise was injected into the training data, teaching the controller to recover from off-path drifts.

Experimental Setup and Datasets

The system was evaluated using the CARLA simulator across several environments (Town01 through Town05), ranging from simple grids to complex multi-lane layouts with roundabouts. A critical component of the evaluation was testing the system's resilience to appearance changes. The researchers conducted tests across diverse conditions, including Clear Noon, Heavy Rain, and Soft Rain Sunset.

Visual examples of the diverse environmental conditions in CARLA, illustrating the significant appearance changes the system must navigate.

Key Results

The performance was measured using a "teach-and-repeat" metric. In simple environments (Town02), the Topo-DLM model achieved a 0.65 mean success rate, whereas basic replay baselines failed completely due to immediate path drift.

In more complex environments like Town03, the Topo-LP (Local Path) model proved superior, achieving a 0.722 success rate. The LP command provided the necessary geometric detail for roundabouts that discrete commands lacked. Notably, the Topo-LP model maintained high performance even under "Soft Rain Sunset" conditions, demonstrating the effectiveness of the Siamese network in handling degraded visual features that typically cause traditional SLAM systems to fail.

Conclusion and Future Work

This research demonstrates that a bio-inspired topometric approach can significantly improve the robustness and efficiency of autonomous route following. By replacing dense metric maps with a series of recognizable waypoints and local paths, the system becomes more resilient to environmental changes.

Future work aims to address the limitations of the current study, specifically by integrating dynamic obstacle avoidance for environments with other vehicles and pedestrians. Additionally, the authors plan to transition from simulator-provided velocity to pure visual odometry, further reducing the system's reliance on internal vehicle sensors.

Summary updated

BibTeX
@inproceedings{cai2020topometric,
  author = {Shaojun Cai and Yingjia Wan},
  title = {{Topometric Imitation Learning For Route Following Under Appearance Change}},
  booktitle = {Workshops @ CVPR},
  year = {2020},
  doi = {10.1109/cvprw50498.2020.00513},
}
Figure