Skip to content

Calibrate Utama field physics from real robot logs - #8

Merged
fred-huang122 merged 2 commits into
masterfrom
calibration/utama-field-20261028
Nov 3, 2026
Merged

fred-huang122 merged 2 commits into
masterfrom
calibration/utama-field-20261028

Conversation

@fred-huang122

@fred-huang122 fred-huang122 commented Nov 3, 2026 •

Copy link
Copy Markdown
Collaborator

grSim's generic robot response and ball rolling differ from the supplied physical Utama field logs. Add an opt-in field configuration with separate models for the two physical robots, fitted from the confirmed 27–28 October runs.

  • Radio 0 and radio 1 have independent forward/left gains and a compact speed-dependent angular response. Other robot models retain unit gains and zero slope. Teleport velocity targets and raw wheel commands keep their existing semantics.
  • The field profile uses 0.20 m/s² rolling deceleration and six robots per side. Geometry/mass remain unmeasured Parsian assumptions. Camera/radio delay, visual noise, kicker charge and dribbler loss are kept separate from these physics parameters.
  • grSim --settings FILE loads an explicit configuration for this run without overwriting personal settings. Explicit XML is validated in the shared loader. Apple bundles include the profiles and test both ordinary and field-profile startup after relocation, including the exact macOS 15 download on macOS 26.
  • grSim_replay --checkpoint-every 1 emits every simulation tick. Snapshot frequency is independent of fixture command cadence. Legacy v1 recordings without response fields retain their original physics.

Physical validation uses the actual ODE solver, original recorded commands and physics-only measured windows. Of 172 release logs, simulator, uncertain 21 October, secondary 29 October, setup-only and unusable sessions are separated before fitting; no raw data, private reports or prototypes are committed. Whole source files are held out for ball fitting; robot gains use held-out command steps within recordings, not independent field days. Observation-path alignment is estimated on fit data and applied equally to both comparisons.

Held-out prediction error Generic model Calibrated
Ball position, pooled RMSE 30.38 mm 18.63 mm
Radio 0 forward / sideways, median RMSE 4.72 / 17.29 mm 3.91 / 9.86 mm
Radio 1 forward / sideways, median RMSE 8.93 / 10.90 mm 7.70 / 5.05 mm
Radio 0 heading, median RMSE 9.79° 2.47°
Radio 1 heading, median RMSE 11.27° 6.95°

Ball scores cover 28 eligible windows from 11 held-out source files. Robot scores cover 27 held-out steps across both robots; angular groups contain three steps each. This is an initial calibration over observed speeds, not validation of all contacts or operating conditions. Missing spin/height/voltage measurements and invalid initial possession prevent reliable kicker/dribbler fitting; those failure snapshots remain separate. Intentional kicks are excluded from dribbler-loss labels.

Validation: the final revision passes all 28 local regressions (103.47 s), both Ubuntu CI builds, and all 27 native regressions on each of Apple Silicon macOS 15 and 26. The relocated macOS 15 package and that exact download on macOS 26 pass ordinary and field-profile startup with OpenGL. CI smoke checks use Apple's software renderer and do not establish hardware GPU frame rates. Standards and spec reviews have zero unresolved findings. Three interleaved 15-second 6 v 6 sustained-pushing trials per profile retain a 0.711 mm maximum shell penetration; the calibrated profile stays at least 2.35 times real time, with no observed performance regression. These headless wall times include startup and snapshots and do not measure GUI FPS. Direct double-precision ODE, fixed timestep and five substeps are retained.

CI: https://github.com/First-Order-RoboCup-SSL/grSim/actions/runs/41570327018

Data source: Utama-Core v2.5.0 field archive, release commit cffe3a2838024504305927d650bd80e105d8e1f2; archive checksums verified.

@fred-huang122
fred-huang122 merged commit a5767e0 into master Nov 3, 2026
5 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant