Automated Mobile Robots (AMRs) are widely used in modern industrial environments to automate material transport and improve operational efficiency. Reliable prediction of transport durations is an important requirement for effective fleet coordination, particularly in dynamic production systems with high traffic density. This tool presents a data-driven approach for estimating AMR travel times based on historical operational data and machine learning methods.
Accurate estimation of transport durations is a key requirement for efficient operation of Automated Mobile Robot (AMR) fleets in semiconductor manufacturing environments. In highly automated facilities, transport processes are closely coupled with production scheduling, meaning that delayed material delivery can directly affect machine utilization and overall throughput.
Traditional estimation methods often rely on static assumptions such as predefined travel speeds or average historical durations. However, these approaches cannot adequately reflect dynamic influences such as temporary congestion, route conflicts, or fluctuating fleet utilization. As the number of robots and transport requests increases, these effects become more pronounced and lead to unreliable travel time predictions.
The motivation behind this tool is therefore to provide a data-driven mechanism for estimating AMR travel times under realistic operating conditions. By incorporating information about the current fleet state and transport context, the approach aims to improve decision-making within fleet management systems and support more stable and predictable intralogistics operations.
The AMR Travel Time Prediction tool is a machine learning–based component designed to estimate the expected duration of transport operations within automated intralogistics systems. The tool processes historical fleet management logs and derives structured datasets that can be used to train predictive models.
Input features describe the operational state of the system at a given point in time. Typical information includes robot positions, transport origins and destinations, intermediate routing points, and contextual indicators related to fleet activity. The output variable corresponds to the measured transport duration of completed tasks.
The prediction problem is formulated as a supervised regression task. Different regression models can be applied, with neural network–based approaches showing strong performance for capturing nonlinear interactions between robots and traffic conditions. Once trained, the model can be integrated into a fleet management environment to provide runtime travel time estimates for ongoing and future transport requests.
The tool supports periodic retraining using newly collected operational data, enabling continuous adaptation to changing layouts, fleet sizes, or production conditions.
A typical implementation for AMR travel time prediction can be structured as a modular machine learning pipeline built in Python. The process starts with loading historical fleet operation logs and transforming them into a structured dataset that represents individual transport tasks, including origin, destination, system state information, and observed travel durations.
df = pd.read_csv("transport_logs.csv")To improve data quality and robustness, preprocessing steps are commonly introduced. These may include filtering implausible transport durations as well as consolidating repeated observations of identical routes by aggregating their measured travel times. Such steps help reduce noise and improve the statistical consistency of the training data.
In addition to tabular route descriptors, more expressive feature representations can be incorporated. A common extension is the use of graph-based embeddings (e.g., Node2Vec) to encode structural relationships between locations in the transport network, enabling the model to capture spatial dependencies beyond simple identifiers.
After feature construction, the dataset is typically converted into a numerical representation and split into training and test subsets to evaluate generalization performance.
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2)For the prediction task, standard supervised regression models can be applied. These range from linear or tree-based models to more expressive approaches such as neural networks. In the case of multilayer perceptrons, the model learns a nonlinear mapping from system state and route descriptors to expected travel time, with the hidden-layer configuration controlling model capacity.
model = MLPRegressor(hidden_layer_sizes=(h1, h2))model.fit(X_train, y_train)Model performance is typically assessed using regression metrics such as Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and the coefficient of determination ($R^2$), which quantify prediction accuracy and variance explanation on unseen data.
mae = mean_absolute_error(y_test, predictions)r2 = r2_score(y_test, predictions)Once trained, the model can be embedded into a runtime prediction interface that accepts a start–destination query (and optional contextual features) and returns an estimated travel time. This enables direct integration into higher-level fleet management or scheduling systems.
predicted_time = model.predict(input_features)The primary purpose of the tool is to support operational decision-making in AMR fleet management systems. Predicted travel times can be used to improve task allocation, route selection, transport prioritization, and scheduling processes.
The approach is particularly suitable for industrial environments characterized by:
Although the methodology was evaluated in semiconductor manufacturing scenarios, it can also be transferred to other automated transport domains such as warehouse logistics, smart factories, or hospital intralogistics systems.
The tool is intended to complement existing fleet management strategies by providing more accurate time estimates than conventional heuristic approaches.
The quality of the predicted travel times strongly depends on the availability and consistency of historical operational data. Incomplete logs, inaccurate timestamps, or missing routing information can reduce model performance and generalization capability.
Prediction accuracy may also decrease in environments that differ significantly from the training conditions. Examples include major layout modifications, changes in fleet behavior, unexpected traffic patterns, or substantially increased robot densities. In highly dynamic systems with unstable traffic conditions, prediction uncertainty can increase considerably.
In addition, the approach focuses on estimating transport durations and does not directly optimize fleet behavior or routing strategies. The generated predictions must therefore be combined with additional fleet management logic for operational deployment.
Finally, retraining and validation are required on a regular basis to ensure that the model remains representative of the current production environment.