Hi ArduPilot community,
I’m Aashrith Bandaru, a CS + Stats student at UIUC, and I’m interested in applying for the SITL Model Generation from Flight Data project. I wanted to engage with the community and get feedback before drafting my proposal.
I’ve been working on simulation pipelines for autonomous systems at UIUC’s BLENDER Lab, including trajectory rollout and counterfactual evaluation, and I also have a strong background in statistical modeling and parameter estimation from time-series data. This made the SITL fidelity problem particularly interesting to me.
My current thinking is to build a grey-box system identification pipeline that parses ArduPilot DataFlash .bin logs to extract synchronized signals (RCOU, IMU, ATT, CTUN), fits key dynamics parameters such as thrust-to-weight ratio, moments of inertia, drag coefficients, and battery model parameters using nonlinear least squares optimization, and outputs a tuned .json frame model and SIM_* .parm file compatible with sim_vehicle.py.
I also plan to support SYSID-mode chirp logs for better parameter identifiability and include validation by comparing simulation vs real behavior on held-out log windows.
A few questions for @NDev:
-
Should I prioritize standard flight logs or SYSID chirp logs as the primary input?
-
Which parameters tend to matter most for improving SITL fidelity in practice?
-
Is there a recommended dataset or reference vehicle to use as a development baseline?
Happy to share a draft proposal soon. Looking forward to contributing!