Change FDM at runtime

[Dobid]

Hello JSBSim Team,

I'm working on a RL-based autopilot for a UAV and I would need to sample random coefficient values let's say +/- 10% around the existing value and apply them at the start of each new training episode. This is for counteracting overfitting of my learned autopilot and to ease the transferability to a real UAV.
However from my knowledge of JSBSim, those values are statically defined in the fdm.xml file and the corresponding properties eg. aero/coefficient/CLalpha are read-only values (from what the print_property_catalog() function returns).

Q: Is it possible to change those FDM coefs are runtime through python code ?

By the way, I'm using the python API of JSBSim v.1.13 installed through conda on a Ubuntu 20.04 computer.

Thanks !

[seanmcleod]

This is what I do.

    <aerodynamics>

        <property value="-0.6">    aero/stability_coefficients/Cm-alpha </property>
        <property value="-27.0">   aero/stability_coefficients/Cm-q </property>
        <property value="-16.0">   aero/stability_coefficients/Cm-alpha-dot </property>
        <property value="1.0">     aero/stability_coefficients/Cm-d-elevator </property>

        <property value="-0.2">    aero/stability_coefficients/Cl-beta </property>
        <property value="-0.4">    aero/stability_coefficients/Cl-p </property>
        <property value="0.09">    aero/stability_coefficients/Cl-r </property>
        <property value="1.0">     aero/stability_coefficients/Cl-d-aileron </property>
        <property value="0.01">    aero/stability_coefficients/Cl-d-rudder </property>

        <property value="0.1">    aero/stability_coefficients/Cn-beta </property>
        <property value="0.00">    aero/stability_coefficients/Cn-p-xx </property>
        <property value="-0.35">   aero/stability_coefficients/Cn-r </property>
        <property value="0.00">    aero/stability_coefficients/Cn-d-aileron-xx </property>
        <property value="-0.2">    aero/stability_coefficients/Cn-d-rudder </property>

        <axis name="PITCH" unit="LBS*FT">
            <function name="aero/coefficient/Cmalpha">
                <description>Pitch_moment_due_to_alpha</description>
                <product>
                    <property>aero/qbar-psf</property>
                    <property>metrics/Sw-sqft</property>
                    <property>metrics/cbarw-ft</property>
                    <property>aero/alpha-rad</property>
                    <property>aero/stability_coefficients/Cm-alpha</property>
                </product>
            </function>
            <function name="aero/coefficient/Cmde">
                <description>Pitch_moment_due_to_elevator</description>
                <product>
                    <property>aero/qbar-psf</property>
                    <property>metrics/Sw-sqft</property>
                    <property>metrics/cbarw-ft</property>
                    <property>fcs/elevator-pos-rad</property>
                    <property>aero/stability_coefficients/Cm-d-elevator</property>
                    <table>
                        <independentVar>velocities/mach</independentVar>
                        <tableData>
                          0             -1.2        
                          2             -0.3        
                        </tableData>
                    </table>
                </product>
            </function>
            <function name="aero/coefficient/Cmq">
                <description>Pitch_moment_due_to_pitch_rate</description>
                <product>
                    <property>aero/qbar-psf</property>
                    <property>metrics/Sw-sqft</property>
                    <property>metrics/cbarw-ft</property>
                    <property>aero/ci2vel</property>
                    <property>velocities/q-aero-rad_sec</property>
                    <property>aero/stability_coefficients/Cm-q</property>
                </product>
            </function>

[Dobid]

Thanks for the answer !

[Dobid]

I'm re-opening this thread, please tell me if I need to open a new discussion about it.
Well, now I have upgraded my lift and drag forces per alpha with table data as such:

<function name="aero/coefficient/CLalpha">
                <description>Lift per radian of alpha</description>
                <product>
                    <property>aero/qbar-area</property>
                    <!-- <property>aero/alpha-rad</property> -->
                    <table>
                        <independentVar lookup="row">aero/alpha-rad</independentVar>
                        <tableData>
                            -1.5714   -0.0028
                            -1.4614   -0.2135
                            -1.3252   -0.4569
                            -1.0581   -0.7448
                            -0.9691   -0.7705
                            -0.7491   -0.6746
                            -0.4925   -0.3937
                            -0.3406   -0.2369
                            -0.2987   -0.3492
                            -0.2725   -0.5412
                            -0.2254   -0.7308
                            -0.1573   -0.5412
                            -0.0682   -0.1596
                            -0.0001    0.1119
                             0.1308    0.6222
                             0.2146    0.8960
                             0.2722    0.5800
                             0.3403    0.2453
                             0.5288    0.4466
                             0.6441    0.5777
                             0.7802    0.7041
                             0.9583    0.7696
                             1.1940    0.6362
                             1.3511    0.4092
                             1.4821    0.1751
                             1.5711    0.0042
                        </tableData>
                    </table>
                    <!-- <value> 4.0203 </value> -->
                </product>
            </function>

And I would like to know how you manage to apply randomization to such data at runtime ?

[seanmcleod]

I'm working on a RL-based autopilot for a UAV and I would need to sample random coefficient values let's say +/- 10% around the existing value and apply them at the start of each new training episode.

So I'm assuming the same applies to your lift and drag data?

I'm also assuming you want to shift the whole lift versus alpha curve say up or down by a maximum of 10%, as opposed to randomizing each entry in the alpha vs CL table?

If so then you can use the same approach I suggested above.

<property value="1.0">    aero/CL-random </property>

<function name="aero/coefficient/CLalpha">
                <description>Lift per radian of alpha</description>
                <product>
                    <property>aero/qbar-area</property>
                    <property>aero/CL-random</property>
                    <table>
                        <independentVar lookup="row">aero/alpha-rad</independentVar>
                        <tableData>
                            -1.5714   -0.0028
                            -1.4614   -0.2135
                            -1.3252   -0.4569
                            -1.0581   -0.7448
                            -0.9691   -0.7705
                            -0.7491   -0.6746
                            -0.4925   -0.3937
                            -0.3406   -0.2369
                            -0.2987   -0.3492
                            -0.2725   -0.5412
                            -0.2254   -0.7308
                            -0.1573   -0.5412
                            -0.0682   -0.1596
                            -0.0001    0.1119
                             0.1308    0.6222
                             0.2146    0.8960
                             0.2722    0.5800
                             0.3403    0.2453
                             0.5288    0.4466
                             0.6441    0.5777
                             0.7802    0.7041
                             0.9583    0.7696
                             1.1940    0.6362
                             1.3511    0.4092
                             1.4821    0.1751
                             1.5711    0.0042
                        </tableData>
                    </table>
                </product>
            </function>

Then at the start of each training episode you simply use a random number generator to generate a value between 0.9 and 1.1 and then set aero/CL-random to this value.

[Dobid]

Oh yes, that could do the job, very clever. Thanks!

[seanmcleod]

@Dobid how about creating a new discussion item as category "Show and tell" to document your RL-based auto-pilot development using JSBSim? I'd be interested in some more details and possibly also asking some questions about it.

[Dobid]

The work I've conducted is part of my PhD thesis and a paper has been submitted about it two weeks ago. I'll keep you updated once I get acceptance (or rejection) notification. Should be coming during the month of April !