Files
SkillFactory/docs/contact_cycle_sim_stress.md
chenlin 9fe9d27865 feat: checkpoint contact-cycle evaluation and RootXY PPO progress
Consolidate next-stage docs, archive superseded narratives, and preserve frozen document inputs. Retain invalid P2 pairs, partial RootXY training, failed attempts and nonselectable checkpoints as explicit limitations.

Include supervised evaluation/training, desktop GPU admission, bounded evidence and atomic-publication monitor repairs with CPU regressions. Ignore local dotenv configuration.

Validation: 1403 passed, 6 skipped for contact-cycle and related contracts/CLI tests. Full-suite legacy ABI and registration tests require Python 3.12/Torch 2.10/pxr; no runtime checks relaxed. No new physical run or accepted skill.
2026-10-15 14:18:17 +08:00

5.0 KiB
Raw Permalink Blame History

Nominal scripted-residual operating-envelope test

This additive non-learning harness runs SkillFactory-Go2FL-ContactCycleSim-v1 with the existing analytic reference plus scripted residual, never a policy model. It changes no controller, reward, termination, functional threshold or asset.

Preregistered development matrix

configs/contact_cycle/simulation_stress_v1.json declares all 12 ordered cases: zero baseline before/after; all-phase sine amplitudes 0.005/0.01/0.025/0.05 rad; constant signed 8-combination residuals 0.025/0.05 rad; tanh-Gaussian normalized latent std 0.1, development seeds 90321/90322; lift/swing-only sine 0.025/0.05 rad. Eight environments run 900 controls each, 100Hz control / 500Hz physics, nominal scenario only, no external disturbance or actuator delay/noise. The environment seed is 90311. There are no omitted cases and no parameter search.

Sine frequency is 0.75Hz; joint offsets are 0, 2π/3, 4π/3; environment offset is eπ/4. Time is control index ×0.01s. Constant signs are the three bits of the environment index (unset=-1, set=+1). Gaussian requests are independent CPU float32 tanh(0.1*randn([900,8,3])), not finite-sample mean-centered. Thus their population mean is zero and typical actual requested std is about 0.005rad. Gating uses the command phase at control start and holds the request across five physics ticks. Only phases 2/3 (lift/swing) enable gated requests. All-phase and gated tests have different domains: gated success cannot establish safe all-phase exploration. Amplitudes are requested joint radians; the unchanged absolute-target projection may shrink effective residuals near bounds.

Evidence and honest failure accounting

Only each environment's FIRST episode is evaluated by the frozen audit_contact_cycle_functional.evaluate and simulation_functional_v1.json. Failed environments can autoreset while remaining environments finish, but later episodes are excluded from all functional rows and saved histories. Physics-rate failure onset (tick, clock, phase, overlapping causes and state) is distinct from the control-boundary termination/reset clock. All six phase sample counts and every functional gate actually evaluated are retained; the frozen checker returns early on incomplete cycles, so missing later gates are not passes. Original physical failures remain enabled.

Each case saves eight initial snapshots, eight summary rows, first-episode reward/action histories, the first four unsuccessful traces, and a representative complete trace when available. Tensor slices are cloned before serialization, not aliased views of a full matrix storage. Trace includes requested/effective residuals, used/next reference, actual target, real contact and fixture/external wrenches, failure flags and actor metadata (unavailable, since no actor). PhysicalSafetyAudit runs every physics tick, including excluded later episodes; its unchanged physical safety guard can abort the matrix.

Top-level harness_verified, baseline_passed, stress case counts and stress episode counts are separate. passed is only the supervisor-compatible harness and baseline gate, explicitly not an overall stress functional pass. Stress failures are results, never a reason to weaken termination or retry for favorable behavior. No learned, robustness, recovery, sim2real, hardware, release, or publication claim follows from this nominal development matrix.

Bounded execution

The fresh maximum-two-hour contract at logs/contact_cycle/autonomous_sim_resume_v1/window.json has SHA256 511d3f98a2c9ec7b78b1cb7f8f72e28ef06de6dcf96ecf2493ac7399d1eefbc8. It is NOT a renewed 12-hour grant. Same-boot CLOCK_BOOTTIME deadline and 120s cleanup margin are binding; wall UTC is diagnostic. The launcher fails closed on boot change, expiry, process conflict or artifact cap (1GiB total window, conservatively including prior window bookkeeping, with 8MiB log/report reserve). It serializes owned launches with a lock, checks the deadline before Kit and every control, checks disk at least once per second and before evidence writes, and limits the supervisor timeout to min(2400s, remaining−130s).

python3 scripts/verify_cycle_sim_stress.py --supervise \
  --output logs/contact_cycle/autonomous_sim_resume_v1/phase01_nominal_residual/attempt01

The stdlib outer launcher removes CONDA_PREFIX/VIRTUAL_ENV and sets PYTHONPATH to $PWD/source/SkillFactory. The sole Kit invocation is via scripts/supervise_sim.py and /home/cen/IsaacLab/isaaclab.sh -p. It writes an immutable preregistration containing the exact plan and source, asset and protected input hashes before launch, checks hashes each case and after shutdown, and retains the supervisor report. All attempt paths must be fresh; failed attempts are preserved. Stop when this matrix is complete.

Results are authoritative in the fresh attempt directory: per-case result.json, matrix_progress.json, final result.json, supervisor.json, and integrity_after.json. Default CPU tests skip Kit and do not count as physics evidence; the real matrix itself supplies PhysX evidence.