Skip to content

CALIBRATION_GUIDE.md

Latest commit

 

History

History
109 lines (74 loc) · 4.21 KB

CALIBRATION_GUIDE.md

File metadata and controls

109 lines (74 loc) · 4.21 KB

Force/Torque Sensor Calibration Guide

Complete guide for calibrating your OpenFT sensor to achieve high-accuracy measurements using a learned calibration matrix.

Why Calibrate?

Geometric approach assumes ideal sensors and perfect tolerances. Reality includes non-linearities, manufacturing variations, cross-coupling, and thermal effects. Calibration provides 5-10x accuracy improvement.

Equipment & Setup

Need: Reference F/T sensor (ATI, Robotiq, OptoForce) OR calibrated weights + lever arms | Rigid mounting | Data logging

Mount: Reference Sensor → Rigid Coupling → OpenFT → Loading Fixture (no compliance!)

Test: Run ft_min_driver.py --port /dev/ttyACM0 --zmq-port 5555 and verify both sensors output data

Data Collection

Planning Dataset

Collect 150-200 samples covering:

  • All 6 axes: ±X, ±Y, ±Z forces and ±Mx, ±My, ±Mz torques
  • Load levels: 20%, 40%, 60%, 80%, 100% of expected range
  • Some combined loads (diagonal forces, force+torque)

Collection Process

Manual method:

  1. Zero both sensors (no load)
  2. Apply known load and wait 2-3 seconds for stability
  3. Record data from both sensors simultaneously
  4. Remove load, wait for zero
  5. Repeat with different loads/directions

Automated script: Subscribe to ZMQ (port 5555) for OpenFT data, read reference sensor, average 10 samples per recording, combine to CSV row with columns: sensor_1..16, force_local_x,y,z, torque_local_x,y,z

Collection Tips

Do: Cover full range, include ±loads on each axis, wait for stability, record zeros periodically Don't: Exceed rated capacity, rush readings, collect only one direction, skip torque if needed

Calculate Calibration Matrix

Verify data:

head -n 5 your_calibration_data.csv  # Check 22 columns, reasonable values

Run calibration:

cd software/calibration
python calc_calibration_matrix.py --csv your_data.csv --output my_calibration.json

Interpreting output:

Loaded 180 valid data points
Matrix rank: 17 (expected: 17)
Condition number: 2.45e+03        ← Should be < 1e4

Calibration quality:
  Force RMSE:  0.0234 N           ← Should be < 1% of full scale
  Torque RMSE: 0.000156 N⋅m      ← Should be < 1% of full scale

Good calibration: RMSE < 1% range, similar error across axes, condition number < 1e4 Poor calibration: RMSE > 5% (collect more diverse data), one axis much worse (check loading), condition number > 1e5 (check sensors)

Using Calibration

Test: Run ft_min_driver.py (terminal 1), then calc_calibration_matrix.py --live --calibration my_calibration.json (terminal 2). Apply known loads to verify.

Integration: Subscribe to ZMQ port 5556, receive JSON with forces and torques keys

Recalibrate: When accuracy degrades, after mechanical changes, temperature range changes, or repairs

Troubleshooting

High RMSE (>5%): Collect more diverse data, check rigid coupling, verify reference sensor, check synchronization, reduce vibration

High condition number (>1e5): Check for damaged sensors, collect data in missing directions, reduce load if sensors saturating

One axis high error: Collect more data in that direction, check reference calibration for that axis, verify no compliance

Drift over time: Allow thermal stabilization, tighten fasteners, avoid external magnetic fields

Calibration Checklist

  • Mount both sensors rigidly (no compliance)
  • Test sensors independently
  • Zero with no load
  • Plan loading matrix (all axes, multiple levels)
  • Collect 150-200 diverse samples
  • Save CSV with correct format
  • Run calc_calibration_matrix.py
  • Verify RMSE < 1% of range
  • Test with known loads
  • Document calibration date/conditions
  • Save with sensor serial number

Why Matrix Calibration Works Better

The calibration matrix learns optimal sensor fusion:

  • Each F/T axis uses weighted combination of all 16 sensors (not just geometric assumptions)
  • Compensates non-linearities in Hall sensor response curves
  • Handles cross-coupling (forces affecting other axes' sensors)
  • Corrects per-sensor variations from manufacturing tolerances

This is why it achieves 5-10x better accuracy than geometric calculations.