跳转至

Physics Results

状态:Pilot / diagnostic;没有 formal learned-method ranking。 target checkout 有 ours 与 PHC-X 的部分 可复核 dev artifacts,但 CoDA/InterMimic/RePHO 的正式 target lifecycle 尚未闭环;HDMI 有 adapter/launcher 代码,但 pinned upstream dependency 和 lifecycle missing;Physics metric 也尚待人选择。

Target-local manipulation-first Full20 pilot(2026-08-12)

20 个 scheduled dev cases 都执行 fresh 200-epoch partial;12 个被开发 reward gate 允许继续到 3000 epoch, 6 个通过三次 frame-0 eval 的 80% terminal opening proxy

Development outcome Cases Denominator
completed full dev training after partial gate 12 20 scheduled
terminal opening proxy pass 6 20 scheduled
terminal opening proxy pass among admitted 6 12 admitted(不可作为主分母)
not admitted / operational failure retained 8 20 scheduled

该 proxy 使用当前开发阈值,只看 terminal/absolute opening;它不是 human-calibrated primary,不证明 pairwise hand--link contact、post-contact causal actuation、all-phase full horizon 或 formal superiority。partial gate 本身还会造成选择偏差,所以 formal training 不得沿用该筛难例规则。

Target-local evidence:

/mnt/cvda_mnt/projects/arthoi4d/output/verification/shared_migration_20260801/
  full20_manipulation_3000_20260812_retry1/full20_analysis.md
  full20_manipulation_3000_20260812_retry1/full20_result_analysis.md
  report.md

这些报告可用于 debug/吞吐/失败分析,不进入 main table。

Target-local interface evidence

Component Verifiable target evidence 结论边界
Ours train/reload/eval output/verification/shared_migration_20260801/ours_d3d_studio_residual_ppo_f/ours_revalidation/ dev lifecycle partial;非 formal split/budget
PHC-X common eval .../phcx_shared_tracking_smoke_d/eval/ 有 rollout/evaluation/run record diagnostic interface,不是 manipulation baseline
CoDA .../baseline_coda/eval/evaluation.json 孤立 dev eval;target 无同目录 full lifecycle provenance
InterMimic backend code 存在 target formal artifact chain missing
RePHO backend code 存在 train→repair/exchange→reload→eval missing
HDMI adapter/launcher/config code exists;target 缺 submodules/HDMI 与 lifecycle missing / time-gated

这些事实与 Physics baselines 的 status table 一致。

Read-only historical author-path evidence

旧文档列出的 PHC-X/CoDA/InterMimic/RePHO 单 case final_author_20260729 产物不在 target,而在:

/DATA/intern/hoi4d/projects/zhanbaiting/ArtHOI4D/
  output/physics_studio_verification/final_author_20260729/

它们是只读旧 checkout 的短训/interface witnesses,只能说明当时曾导出 common rollout;不能写成当前 target author reproduction、formal adapter parity 或效果排名。若论文需引用,必须生成 source checkout、 commit/diff、input/output/config/evaluator hashes 的 evidence manifest,或在 target 重新运行。

当前不能写成结论

  • 不能声称 formal Physics primary 已冻结;DEV YAML thresholds 没有人体盲评/测量学依据。
  • 不能称 proximity + marginal force 是 verified same-link pair contact。
  • 不能称当前 case bootstrap 为 CAD/object clustered bootstrap。
  • 不能比较 CoDA/InterMimic/RePHO/HDMI/Ours 的正式性能。
  • 不能以 6/20 terminal proxy 推断论文级 task success 或否定未来 method。

正式结果需要 matched Studio lifecycle、每个 scheduled case 一次 canonical rollout、scheduled ITT、true CAD clusters、valid-prefix coverage、 rad/m separation、signed metric contract 和 private test。设计入口: Physics metrics · Evaluation protocol · 28-day plan