03. Batch Training

Batch Training Process

Release 0.27.0

The batch training process converts data collected during an online session into training datasets and uses them for batch learning and reporting.

Run

After collecting data with RobotExecutor or Wheelly, run:

bin\batch

To stop the batch process, delete the .running file.

Workflow

flowchart TD
    A[Online Session] --> B[Inference Data]
    B --> C[CreateDatasets]
    C --> D[Training Data]
    D --> E[BatchTraining]
    E --> F[KPIs / Report Data]
    F --> G[Report Document]
    G --> H{.running exists?}
    H -->|Yes| E
    H -->|No| I[End]

Loading

The process repeats while .running exists.

CreateDatasets

CreateDatasets converts the inference file into training data.

The generated files are stored in %TEMP%:

• canMoveStates.bin
• map.bin
• move.bin
• rewards.bin
• sensor.bin
• sensorAction.bin

Command:

call bin\runnit.cmd org.mmarini.wheelly.apps.CreateDatasets -t %TEMP% %INFERENCE%

The default configuration file is batch.yml.

BatchTraining

BatchTraining uses the datasets created by CreateDatasets to perform batch learning.

Command:

call bin\runnit.cmd org.mmarini.wheelly.apps.BatchTraining -u -k %KPIS% %TEMP%

Main parameters:

Parameter	Description
%TEMP%	Training dataset path
%KPIS%	KPI output directory
-u	Opens the performance page

Reporting

After training, the KPI and reward data are processed:

call bin\report.cmd -p -r %TEMP%/rewards.bin %KPIS% %CSVS%

The final report is generated by GNU Octave:

cd octave
"C:\Program Files\GNU Octave\Octave-7.2.0\octave-launch" --no-gui dlReport.m
cd ..

Training Loop

The batch script repeats the training and reporting steps:

BatchTraining
     ↓
Report Generation
     ↓
Next Iteration

The .running file controls the loop:

• File exists → continue training.
• File deleted → stop after the current iteration.

Summary

The complete process is:

Online Data → CreateDatasets → BatchTraining → KPI/Report Generation → Next Iteration

This provides an automated way to repeatedly train and evaluate the model using data collected from online sessions.