Skip to content

Preserve calibrated motion with adaptive collision stepping - #9

Merged
fred-huang122 merged 8 commits into
masterfrom
calibration/utama-turning-response
Nov 4, 2026
Merged

fred-huang122 merged 8 commits into
masterfrom
calibration/utama-turning-response

Conversation

@fred-huang122

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

Copy link
Copy Markdown
Collaborator

Recorded turns under-predict the response of Utama radio 1, and a 6 v 6 pushing/turning case exceeds the existing 2 mm shell-overlap bound. Refine the fitted radio 1 angular limits and use a shared adaptive integration schedule to preserve measured motion while resolving first impacts.

The adaptive policy uses two evenly spaced steps per fixed tick for settled, distant motion; at least five for startup, placement/reset, acceleration, turning, sliding or flight; and finer subdivisions for predicted robot, ball and wall impacts. Every body uses the same fixed simulation clock. The direct double-precision solver, collision geometry, contact material parameters and 2 mm bound are unchanged. Modern commands, legacy replacements, GUI placement/formation and automatic turnover resets all restart settling.

Replay v2 records the explicit integration policy. Historical v1 retains fixed-five physics; the prior near-shell-six policy also remains explicit. Controlled fixed-count replay supports the requested 1/2/3/5 comparisons. Unknown policies and incompatible version/count combinations are rejected.

Radio 1 angular speedup/braking limits change from 50 to 125 while its velocity gains and radio 0 remain unchanged. These are fitted simulator control settings, not measured hardware acceleration maxima. The original three withheld turns improve from median 6.95° to 3.22° heading RMSE; the expanded 13 turns improve from 2.98° to 1.83°. Stopping remains a disclosed trade-off: capture-clock median 2.17° → 2.00°, receive-clock median 2.09° → 2.14°.

Validation uses 84 physical movement events and 21 stopping phases from confirmed 27–28 October logs, plus 161 rolling-ball windows (28 held out from 11 source files). SIM, uncertain-origin, incomplete and contact-sensitive records remain separate. Parameters, initial poses, observation delays and original transmitted command cadence are frozen across integration comparisons. Actual capped commands are replayed; these data do not establish onboard maximum speeds. Weak/no-fire and off-centre kicks remain interpretation context; kicker physics is not retuned. This is development validation on the existing archive, not newly collected independent evidence.

Integration Radio 0 median turn RMSE Radio 1 median turn RMSE
Fixed 1 1.177° 3.018°
Fixed 2 1.216° 1.953°
Fixed 3 1.115° 2.000°
Fixed 5 1.100° 1.828°
Adaptive, minimum 2 1.100° 1.828°

Adaptive changes each movement's driving-axis RMSE by at most 0.033 mm versus fixed five; turning and stopping predictions are numerically unchanged at the reported precision. Held-out ball pooled RMSE is 18.628 mm with fixed five and 18.633 mm adaptive. No parameters were refitted to hide integration differences.

Three interleaved 15-second, 6 v 6 trials with three ODE helpers:

  • Separated driving: median wall time 1.392 s fixed-five → 0.679 s adaptive; 4,500 → 2,043 integration steps, zero shell contacts and minimum sampled robot separation 0.70 m.
  • Sustained pushing: 6.749 → 7.027 s; adaptive peak overlap 0.888 mm.
  • Pushing with ±2 rad/s turns: 2.317 → 2.846 s; overlap 2.232 → 0.674 mm.
  • Pushing with ±4 rad/s turns: fixed-five peak 3.108 mm, prior near-six peak 2.519 mm, adaptive 0.700 mm; adaptive median wall time 2.658 s.

Adaptive is faster in separated motion and can cost more during dense impacts; every tested adaptive case exceeds real-time throughput, including the slowest sustained pushing case at 2.13×. Global fixed 1/2/3 are rejected: turning crowds exceed 2 mm. These are physics-only CPU timings, not rendered-FPS claims; user background workloads were left running. Synthetic impacts do not replace instrumented physical collision measurements.

Permanent regressions cover the ±2/±4 cases, front/back ball-robot shots, ball/robot-wall impacts, placement settling, exact serial repetition (1e-12), and the existing three-second helper comparison (1e-6). No fifteen-second cross-helper bitwise identity is claimed. All 28 local regressions pass. Standards and Spec review: zero unresolved findings after fixes. Raw logs, generated snapshots, private scripts and analysis remain outside Git.

CI: https://github.com/First-Order-RoboCup-SSL/grSim/actions/runs/41755171142
Validation complete: both Ubuntu 22.04/24.04 pass 28/28 tests; native Apple Silicon macOS 15/26 pass 27/27. The exact macOS 15 artifact (ID 13288466473, SHA256 8cfe5b61eccd2d8fabc95f491595a5b98d05f67a971e70d4c6502aaf4117936b) also launches on macOS 26 with both generic and field profiles. One targeted macOS 26 job retry was needed after an initial application-smoke timeout; the unchanged test passed in 6.23 s against the same 20 s limit. No checks or tolerances were weakened.

@fred-huang122 fred-huang122 changed the title Improve Utama radio 1 turning response Improve turning calibration and first-impact contact accuracy Nov 3, 2026
@fred-huang122
fred-huang122 marked this pull request as draft November 3, 2026 18:22
@fred-huang122 fred-huang122 changed the title Improve turning calibration and first-impact contact accuracy Preserve calibrated motion with adaptive collision stepping Nov 4, 2026
@fred-huang122
fred-huang122 marked this pull request as ready for review November 4, 2026 01:56
@fred-huang122
fred-huang122 merged commit 8e7c3be into master Nov 4, 2026
9 of 10 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