Abstract
Mobile robotic control models typically rely on pre-defined Impact Factors (IFs) to translate sensor data into movement. In dynamic, "open" environments, these fixed coefficients often fail to adapt to changing physical conditions. This paper proposes a clustering-based framework that automatically adjusts IFs at runtime. By partitioning environmental data using density-based clustering (DBSCAN) and learning specialized adjustment models through weighted regression, the system enables robots to maintain stability and precision. Experimental results in path-following and object-following scenarios demonstrate that this approach achieves faster response times and lower deviation compared to Deep Q-Network (DQN) and Model Predictive Control (MPC) methods.
Introduction
Traditional robotic controllers, such as PID or Model Predictive Control (MPC), are designed with fixed parameters—Impact Factors (IFs)—that are tuned for specific, predictable environments. However, real-world "open environments" introduce unpredictable stressors like varying ground friction, slopes, or obstacles. In these scenarios, a single set of IFs is insufficient. While machine learning offers potential for dynamic adaptation, standard reinforcement learning techniques often suffer from "mutual interference," where data from disparate environmental conditions degrades the model's overall performance.
The Self-Adjusting Framework
The proposed solution addresses these challenges through a two-stage process: offline learning and online adjustment. The core of this system is the Learning-Based Adjustment Model (LBAM).
Offline Learning: Environment Partitioning
Instead of training a monolithic model, the framework uses DBSCAN (Density-Based Spatial Clustering of Applications with Noise) to group historical data into "sub-environments." These clusters are formed based on the physical effect of the environment on the robot (e.g., velocity deviation) rather than visual classification. For each cluster, Weighted Least Squares (WLS) regression is applied to map sensor inputs to optimal IFs, prioritizing high-quality data points.
Online Adjustment: Sensitivity-Based Updates
At runtime, the robot identifies the current environmental cluster and calculates the partial derivatives of the learned models. To prevent erratic "trembling" caused by simultaneous parameter changes (the coupling effect), the system identifies and updates only the most sensitive IF—the parameter that will yield the greatest improvement in control performance.
Experimental Evaluation
The framework was validated through both simulation and physical experiments:
- Simulation: Using the V-rep simulator with a motorbike model, the system was tested on paths with varying complexity and ground friction. While MPC and DQN models struggled with balance in complex scenarios, the proposed approach maintained a low deviation from the target path.
- Physical Testing: A Turtlebot platform was deployed in a challenging "mud land" environment. The LBAM-optimized controller allowed the robot to navigate pits and slippery surfaces that caused standard controllers to fail.
Key Insights
- Effect-Based Clustering: By clustering based on the robot's physical response rather than environmental labels, the system generalizes better to novel but physically similar terrains.
- Computational Efficiency: The online adjustment phase has a complexity of $O(m)$, where $m$ is the number of clusters. This makes it significantly faster and more suitable for real-time robotic applications than deep reinforcement learning.
- Stability through Sensitivity: Updating only the most sensitive parameter effectively mitigates the interference between control factors, leading to smoother physical actions.
Conclusion
This research presents a robust framework for self-adjusting control parameters in mobile robotics. By combining unsupervised clustering with targeted regression, the system overcomes the limitations of fixed-parameter models in open environments. Future work will focus on transitioning to a fully online model capable of creating new environmental clusters and updating regression parameters dynamically without an offline training phase.
Summary updated