Replace assumptions with measurements.
Run the same experiment several times on your intended floor. Avoid wheel slip when measuring wheel radius. Effective turning track depends on skid and floor grip.
Calibration evidence
Import your existing calibration first to retain its other values. Otherwise unmeasured parameters remain the app defaults.
No measurements calculated yet.
In the simulator: Physics → Import calibration. Estimates include their measurements and assumptions. No stall-torque or inertia measurement is invented.
How close is your simulation?
Import a real run and simulated data.csv. Select equivalent columns in SI units. Initial translation/yaw offsets are removed; time is relative to each run's first valid sample. No extrapolation is used.
Real robot CSV
Column mapping & units
Simulation CSV
Column mapping & units
● Real / reference● Simulation
Inspect the exported bytes.
Load a dataset ZIP to check timestamps, image references, actual image decoding, blank frames, CSV values and SHA-256 file checksums. The ZIP stays on your computer.
Choose an exported dataset ZIP.
Uniform or clipped images are review warnings: a real empty/dark scene can also produce them. Passing these checks verifies file consistency, not real-world accuracy.
Inspect a recorded drive.
Open drive-replay.json or replays.json from your dataset. Re-execute 120 Hz motor inputs and inspect recorded trajectory and encoder drift. Recorded scene/contact corrections are applied explicitly.
This is drive replay plus recorded corrections, not an independent reconstruction of all dynamic objects or a full 6-DOF physics validation.
Keep related recordings together.
Import metadata.json files from independent recordings or full dataset ZIPs. All runs sharing an environment stay in the same split by default. Never randomly split neighbouring frames.