Scholarlink

This paper proposes a SLAM localization method based on deep reinforcement learning for adaptive covariance matrix estimation in factor graphs, to address the challenge of environment-aware covariance adjustment in pose factor graphs for multi-sensor fusion localization involving vision, LiDAR, and IMU. A pose factor graph fusion module utilizes real-time covariance estimation results to fuse pose estimates from two odometry sources using a sliding window approach. A weight estimation module, composed of a feature extraction layer capable of processing multimodal sensor data and a decision network with temporal learning capabilities, is designed to predict adaptive factor weights. Considering the non-differentiable nature of factor graph optimization, a reinforcement learning-based training method is proposed to enable the network to learn an effective weight prediction policy. Compared to fixed-covariance methods, the proposed approach improves RMSE by 10% on training and 3% on testing sequences of the KITTI dataset.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Self-adaptive Factor Graph Weight Tuning SLAM Method Based on Reinforcement-Learning Method

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Wenhui Wan,
  • Yingxiang Xu,
  • Man Peng,
  • Kaichang Di

摘要

This paper proposes a SLAM localization method based on deep reinforcement learning for adaptive covariance matrix estimation in factor graphs, to address the challenge of environment-aware covariance adjustment in pose factor graphs for multi-sensor fusion localization involving vision, LiDAR, and IMU. A pose factor graph fusion module utilizes real-time covariance estimation results to fuse pose estimates from two odometry sources using a sliding window approach. A weight estimation module, composed of a feature extraction layer capable of processing multimodal sensor data and a decision network with temporal learning capabilities, is designed to predict adaptive factor weights. Considering the non-differentiable nature of factor graph optimization, a reinforcement learning-based training method is proposed to enable the network to learn an effective weight prediction policy. Compared to fixed-covariance methods, the proposed approach improves RMSE by 10% on training and 3% on testing sequences of the KITTI dataset.