DOA: A Degeneracy Optimization Agent with Adaptive Pose Compensation Capability based on Deep Reinforcement Learning

Yanbin Li, Canran Xiao, Hongyang He, Shenghai Yuan, Zong Ke, Jiajie Yu, Zixiong Qin, Zhiguo Zhang, Wenzheng Chi, Wei ZhangarXiv 2025 · DOI · arXiv

Abstract

Simultaneous Localization and Mapping (SLAM) is a foundational technology for autonomous mobile robots. However, 2D LiDAR-based SLAM frequently encounters the "long corridor" problem, where featureless environments lead to geometric degeneracy and significant localization drift. This paper introduces the Degeneracy Optimization Agent (DOA), a deep reinforcement learning-based framework designed to provide adaptive pose compensation. By utilizing Proximal Policy Optimization (PPO) and a Multi-Head Attention architecture, the DOA dynamically adjusts the weighting between a robot's internal motion model and its LiDAR observations. Experimental results in both simulation and real-world settings demonstrate that DOA can reduce Absolute Trajectory Error (ATE) by up to 95.2% compared to standard GMapping, while maintaining real-time performance.

Introduction: The Challenge of Degeneracy

In autonomous navigation, 2D LiDAR-based SLAM algorithms often struggle in degenerate environments—spaces like long, featureless corridors or expansive empty halls. In these scenarios, the LiDAR sensor perceives identical geometry (e.g., parallel flat walls) regardless of the robot's position along the corridor's axis. This lack of longitudinal geometric constraints results in "degeneracy," where the SLAM algorithm cannot distinguish its position, leading to massive localization drift and distorted environmental maps. Traditional solutions often rely on rigid, hand-tuned thresholds that fail to generalize across diverse architectural layouts.

The DOA Framework

The proposed Degeneracy Optimization Agent (DOA) acts as an intelligent supervisor for the SLAM process. It is trained using Deep Reinforcement Learning (DRL), specifically the Proximal Policy Optimization (PPO) algorithm, to observe and correct the SLAM state in real-time.

Overview of the DOA framework, illustrating the fusion of motion and observation constraints guided by the degeneracy factor.

Dynamic Compensation and Fusion

The core of the DOA is its ability to output a degeneracy factor ($a \in (0, 1)$). This factor determines the relative trust placed in the robot's sensors:

  • Feature-Rich Environments: When the environment provides clear geometric landmarks, $a$ remains low, and the system prioritizes LiDAR observations ($z_m$).
  • Degenerate Environments: Upon detecting degeneracy, $a$ increases, signaling the system to shift the pose estimate toward the more reliable internal motion model ($u_m$).

Instead of a binary switch, DOA performs a linear interpolation between the centroids of the motion and observation distributions, ensuring a smooth and adaptive pose correction.

Architecture and Training Strategy

To effectively detect the subtle signs of degeneracy, the DOA employs a sophisticated neural network architecture.

The DOA neural network architecture featuring a Multi-Head Attention mechanism to analyze particle distributions.

Multi-Head Attention

The network treats the "swarm" of hypothetical robot positions (particles) as a set of related data points. By applying a Multi-Head Attention mechanism, the agent can simultaneously analyze the entire distribution. This allows it to identify "distribution stretching"—a mathematical indicator of SLAM failure—more effectively than traditional statistical methods.

Transfer Learning

Training a DRL agent directly in a featureless environment is challenging due to sparse rewards. The authors utilize a two-stage training process:

  1. Pre-training: The DOA is first trained in a "normal," feature-rich environment to learn the characteristics of accurate localization.
  2. Fine-tuning: The "backbone" of the network is frozen, and the agent is fine-tuned in degenerate scenes. This transfer learning approach ensures the agent understands "good" localization before attempting to solve the "long corridor" problem.

Experimental Evaluation

The DOA was evaluated across diverse settings:

  • Simulation: Using Gazebo, the authors tested the agent in standard rooms, degenerate corridors, and complex circular paths.
  • Real-World: Tests involved a wheeled robot in long straight corridors and empty outdoor spaces.
  • Baselines: DOA was compared against industry standards (GMapping, Cartographer, Hector SLAM) and recent state-of-the-art (SOTA) learning-based methods.

Key Results

  • Accuracy: In the most challenging simulated corridor, DOA reduced the ATE of GMapping by 95.2%.
  • Robustness: Compared to other learning-based methods, DOA improved localization success rates by over 30% in several scenes. In real-world tests, while Cartographer suffered from "movement stagnation" in corridors, DOA maintained a near-perfect trajectory.
  • Efficiency: With an inference time of 15-18ms, DOA operates well within the standard 170ms update cycle of 2D-SLAM, making it suitable for real-time deployment on mobile robots.

Conclusion and Future Work

The DOA framework provides a robust, learning-based solution to the long-standing problem of geometric degeneracy in 2D-SLAM. By intelligently fusing sensor data through an attention-based DRL agent, it achieves superior localization accuracy in environments that typically cause standard algorithms to fail.

Future research will focus on handling variable particle counts through distribution embeddings and extending the DOA logic to 3D LiDAR and Visual-Inertial Odometry (VIO) systems. Additionally, optimizing the computational load for low-power embedded systems remains a key objective.

Summary updated

Abstract

Particle filter-based 2D-SLAM is widely used in indoor localization tasks due to its efficiency. However, indoor environments such as long straight corridors can cause severe degeneracy problems in SLAM. In this paper, we use Proximal Policy Optimization (PPO) to train an adaptive degeneracy optimization agent (DOA) to address degeneracy problem. We propose a systematic methodology to address three critical challenges in traditional supervised learning frameworks: (1) data acquisition bottlenecks in degenerate dataset, (2) inherent quality deterioration of training samples, and (3) ambiguity in annotation protocol design. We design a specialized reward function to guide the agent in developing perception capabilities for degenerate environments. Using the output degeneracy factor as a reference weight, the agent can dynamically adjust the contribution of different sensors to pose optimization. Specifically, the observation distribution is shifted towards the motion model distribution, with the step size determined by a linear interpolation formula related to the degeneracy factor. In addition, we employ a transfer learning module to endow the agent with generalization capabilities across different environments and address the inefficiency of training in degenerate environments. Finally, we conduct ablation studies to demonstrate the rationality of our model design and the role of transfer learning. We also compare the proposed DOA with SOTA methods to prove its superior degeneracy detection and optimization capabilities across various environments.
BibTeX
@misc{li2025doa,
  author = {Yanbin Li and Canran Xiao and Hongyang He and Shenghai Yuan and Zong Ke and Jiajie Yu and Zixiong Qin and Zhiguo Zhang and Wenzheng Chi and Wei Zhang},
  title = {{DOA: A Degeneracy Optimization Agent with Adaptive Pose Compensation Capability based on Deep Reinforcement Learning}},
  year = {2025},
  eprint = {2507.19742},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2507.19742},
}
Figure