BEVNav: Robot Autonomous Navigation via Spatial-Temporal Contrastive Learning in Bird's-Eye View

Jiahao Jiang, Yuxiang Yang, Yingqi Deng, Chenlong Ma, Jing Zhang

Abstract

Autonomous navigation in crowded, map-less environments remains a significant challenge for mobile robots. Traditional methods often rely on 2D depth images, which struggle to capture complex 3D spatial relationships and dynamic obstacle behaviors. This paper introduces BEVNav, a novel navigation framework that leverages Bird’s-Eye View (BEV) representations of 3D point clouds. By integrating spatial and temporal contrastive learning as auxiliary tasks within a Deep Reinforcement Learning (DRL) framework, BEVNav enables robots to learn robust environmental representations and predict scene dynamics. Experimental results in high-fidelity simulations demonstrate that BEVNav achieves state-of-the-art success rates and path efficiency in dense, dynamic crowds.

Introduction

Navigating through a busy pedestrian area requires a robot to understand not just where obstacles are, but how they move and how its own actions will affect future states. Most current systems utilize front-view depth cameras. However, these 2D projections often lose critical geometric information and make it computationally expensive to map observations to 3D control actions.

BEVNav addresses these limitations by shifting the perspective to a Bird’s-Eye View. This representation preserves spatial relationships and simplifies the action-mapping process. To enhance the quality of the learned features, the framework employs self-supervised contrastive learning, allowing the robot to learn from unlabeled interaction data.

Methodology

The BEVNav architecture is designed to process 3D point clouds and output optimal navigation commands (linear and angular velocity). The system consists of three main pillars: efficient feature extraction, spatial consistency, and temporal prediction.

Overview of the BEVNav framework illustrating the integration of contrastive learning modules with the reinforcement learning pipeline.

Sparse-Dense BEV Network

The core of the perception system is a Sparse-Dense BEV Network. Raw 3D point clouds are first voxelized into "pillars." Height information within each pillar is compressed to create a 2D BEV map. This map is processed using sparse convolutions to handle the inherent sparsity of point clouds, followed by dense convolutions to extract high-level latent features. This approach effectively filters out irrelevant vertical noise while focusing on navigable space.

Contrastive Learning Modules

To improve the robustness and predictive capabilities of the encoder, two auxiliary contrastive learning tasks are introduced:

  1. Spatial Contrastive Learning (SCL): Inspired by SimSiam, SCL trains the encoder to produce consistent features for different augmentations of the same point cloud. This ensures the robot's perception remains stable despite sensor noise or slight viewpoint changes.
  2. Temporal Contrastive Learning (TCL): TCL aligns the robot's current observation and planned action sequence with future observations. By predicting future states, the robot learns the underlying dynamics of the environment and the consequences of its actions.
Detailed breakdown of (a) BEV Feature Extraction, (b) Spatial Contrastive Learning, and (c) Temporal Contrastive Learning modules.

Reinforcement Learning

The extracted features are fed into a Soft Actor-Critic (SAC) algorithm. The reward function is multi-objective, incentivizing the robot to reach the goal, maintain a safe distance from obstacles, and move efficiently.

Experimental Results

BEVNav was evaluated in the Gazebo simulator using the PEDSIM library to model realistic pedestrian dynamics. Testing occurred in two environments: "Square-World" for training and "Lobby-World" for complex testing.

  • Navigation Performance: In a cluttered lobby environment without pedestrians, BEVNav achieved a 94% success rate, outperforming depth-based methods (87%).
  • Crowd Robustness: In scenarios with 20 moving pedestrians, BEVNav maintained a 75% success rate, significantly higher than PointNet-based (55%) and depth-based (68%) alternatives.
  • Path Efficiency: BEVNav consistently achieved higher Success Path Length (SPL) scores, indicating more direct and efficient routes to the goal.

Conclusion

BEVNav demonstrates that Bird’s-Eye View representations, combined with spatial-temporal contrastive learning, significantly enhance robot navigation in dynamic, crowded environments. While the system shows strong performance in simulation, future work will focus on addressing sensor blind spots and transitioning the model to real-world physical platforms with multi-modal sensor suites.

Summary updated

BibTeX
@article{jiang2024bevnav,
  author = {Jiahao Jiang and Yuxiang Yang and Yingqi Deng and Chenlong Ma and Jing Zhang},
  title = {{BEVNav: Robot Autonomous Navigation via Spatial-Temporal Contrastive Learning in Bird's-Eye View}},
  journal = {IEEE Robotics Autom. Lett.},
  year = {2024},
  doi = {10.1109/lra.2024.3482190},
}
Figure