| name | sim2real-calibration |
| description | Sim-to-real calibration skill for robot-dog virtual prototypes. Use this skill
when comparing simulation logs against physical prototype logs, estimating
friction/mass/latency/torque scale/contact parameter corrections, reducing
sim-real gap, or generating parameter updates for MuJoCo/PyBullet/digital twin.
|
sim2real-calibration
This skill compares simulation metrics with measured prototype metrics and
proposes conservative parameter updates. The MVP is file-based and does not
connect to hardware. It helps reduce the sim-real gap after a physical prototype
exists.
When To Use
Use this skill for:
- Comparing simulated and measured speed, slip, torque, roll/pitch, and latency.
- Producing parameter update suggestions for simulation/MuJoCo.
- Blocking digital-twin trust when sim-real error is too high.
- Creating calibration reports after physical testing.
Workflow
- Read
<project>/calibration_dataset.yaml.
- Compare
simulation and real metrics against tolerances.
- Estimate conservative parameter updates.
- Write calibration report artifacts into
<project>/reports/.
Commands
python skills/sim2real-calibration/scripts/compare_logs.py skills/sim2real-calibration/examples/quadruped_mvp
python skills/sim2real-calibration/scripts/propose_parameter_update.py skills/sim2real-calibration/examples/quadruped_mvp
Rules
- Never claim calibrated parameters are validated without a follow-up simulation and physical re-test.
- Treat missing real logs or excessive error in required metrics as blockers.
- Keep updates conservative and bounded.
- Do not import sibling subskill code. Read files only.
References
references/input-contract.md for calibration_dataset.yaml fields.
references/metrics.md for comparison metrics.
references/update-rules.md for parameter update heuristics.
references/output-contract.md for report fields.
references/examples.md for sample project layout.