exp: 轮次38(补)——steps=4 末态误差0.135mm最优但仍被拒80次/题；失败分类证明缺口是走一半被冻住(63/109)
This commit is contained in:
@@ -0,0 +1,73 @@
|
||||
{
|
||||
"config": {
|
||||
"ckpt": "ml/runs/base_dag2_dag1/best.pt",
|
||||
"samples": "ml/data/samples",
|
||||
"tasks": null,
|
||||
"split": "val",
|
||||
"n": 256,
|
||||
"K": 8,
|
||||
"commit": 2,
|
||||
"fb": 2.2,
|
||||
"ki": 0.0,
|
||||
"fb_scale": 0.0,
|
||||
"fb_near": 0.2,
|
||||
"fb_fallback": 0.0,
|
||||
"kdecay": 0.9,
|
||||
"kimax": 0.5,
|
||||
"fb_every": 1,
|
||||
"safety_try": 0,
|
||||
"sel_w": 0.1,
|
||||
"dls_safety": 1,
|
||||
"fb_clear": 0.0,
|
||||
"fb_cmin": 0.3,
|
||||
"sub_refresh": 0,
|
||||
"fb_block_gain": 0.5,
|
||||
"dump": "ml/runs/round36_steps4_cap1_rows.json",
|
||||
"out": "docs/experiment_round36_steps4_cap1.json",
|
||||
"trace": null,
|
||||
"select": "endpoint",
|
||||
"replan": "empty",
|
||||
"safety": 1,
|
||||
"partial": 1,
|
||||
"inc_cap": 1,
|
||||
"gap_margin": 0.0,
|
||||
"cert_batch": 0,
|
||||
"steps": 4,
|
||||
"backoff": 0,
|
||||
"max_sec": null,
|
||||
"seed": 0,
|
||||
"dls": "analytic",
|
||||
"device": "cuda"
|
||||
},
|
||||
"batch_size": 256,
|
||||
"batch_seed_rule": "same seed per batch",
|
||||
"checkpoint_sha256": "f7f686fd8ee92fb6948fbcc4ee954636065ad87bb54f661c4f8f575bc16c2ce1",
|
||||
"tasks_sha256": "249bd93457825cda98c59e64d9a096d3473c69aeb3553de1dea6d9e56be71772",
|
||||
"kin_fp": "2ebcbf17",
|
||||
"act_scale": 1.0,
|
||||
"metrics": {
|
||||
"n": 256,
|
||||
"n_ref": 256,
|
||||
"success": 0.40625,
|
||||
"collision_free_success": 0.40625,
|
||||
"time_ok_1p3": 0.3046875,
|
||||
"collision_rate": 0.0,
|
||||
"collision_steps_mean": 0.0,
|
||||
"travel_ratio_median": 1.8817013502120972,
|
||||
"time_mean_s": 1.9080770015716553,
|
||||
"brakes_mean": 79.65234375,
|
||||
"start_fail": 0.0,
|
||||
"goal_ratio_2s": 0.3844277262687683,
|
||||
"err_fail_med": 342.1692810058594,
|
||||
"err_success_med": 0.1346309632062912,
|
||||
"dls_share": 0.4609375,
|
||||
"time_budget": "teacher×2(≥3 s)",
|
||||
"speed_ratio_p50": 1.0000016689300537,
|
||||
"speed_ratio_p95": 1.0000053644180298,
|
||||
"speed_ratio_max": 1.0000101327896118,
|
||||
"over_vmax_rate": 0.0,
|
||||
"accel_ratio_p95": 0.914295494556427,
|
||||
"accel_ratio_max": 1.0000005960464478
|
||||
},
|
||||
"simulation_wall_s": 347.79689209999924
|
||||
}
|
||||
@@ -0,0 +1,19 @@
|
||||
[
|
||||
{
|
||||
"label": "steps=1 cap=0",
|
||||
"n": 256,
|
||||
"arrived": 147,
|
||||
"failed": 109,
|
||||
"buckets": {
|
||||
"姿态没到位(位置<5mm)": 3,
|
||||
"位置 5~20mm": 5,
|
||||
"位置 20~100mm": 15,
|
||||
"位置 >100mm 且刹车>30(被挡)": 63,
|
||||
"位置 >100mm 且刹车<=30(方向/绕圈)": 23,
|
||||
"起步失败": 0
|
||||
},
|
||||
"failed_e_p50": 255.333598802122,
|
||||
"failed_oe_p50": 11.016554443385107,
|
||||
"failed_still_share_p50": 0.3722222149372101
|
||||
}
|
||||
]
|
||||
+16
-1
@@ -34,8 +34,23 @@
|
||||
| steps=2 cap=0 | 0.4258 | 0.3555 | **86.31** | 1.67 s | **1.571** | 0.243 mm | 0.453 |
|
||||
| steps=1 cap=1（=轮次33 cap） | 0.5234 | 0.2656 | 19.55 | 2.49 s | 2.616 | 0.218 mm | 0.613 |
|
||||
| steps=2 cap=1 | 0.3945 | 0.2969 | 86.16 | 1.93 s | 1.844 | 0.293 mm | 0.445 |
|
||||
| steps=4 cap=1 | 0.4062 | 0.3047 | 79.65 | 1.91 s | 1.882 | **0.135 mm** | 0.461 |
|
||||
|
||||
（`docs/experiment_round36_steps2_cap{0,1}.json`；steps=4/8 同批续跑）
|
||||
（`docs/experiment_round36_steps2_cap{0,1}.json`、`docs/experiment_round36_steps4_cap1.json`；steps=8 同批续跑）
|
||||
段数再多并不是越差：steps=4 比 steps=2 到达与时限都略升，成功题末态误差 0.293 → **0.135 mm**（比一步口径的 0.218 还好）——但刹车仍近 80/题，到达率被同一道门卡住。
|
||||
- **失败分类**（不占卡：读轮次 33 base 的 256 题轨迹转储，CPU 重算末态位姿误差，`docs/experiment_round38_fail_taxonomy.json`）。147 到达 / 109 未到，未到题按"末态离目标多远 + 被拒多少次"分桶：
|
||||
|
||||
| 桶 | 题数 |
|
||||
|---|---|
|
||||
| 位置 >100 mm 且刹车 >30 次（走一半被冻住） | **63** |
|
||||
| 位置 >100 mm 但刹车 ≤30 次（方向错 / 绕圈） | 23 |
|
||||
| 位置 20~100 mm | 15 |
|
||||
| 位置 5~20 mm | 5 |
|
||||
| 位置 <5 mm 但姿态没到位 | 3 |
|
||||
| 起步失败 | 0 |
|
||||
|
||||
未到组中位数：位置误差 255 mm、姿态误差 11.0°、**活动窗内 37% 的步是零运动**。
|
||||
⇒ 到达率的缺口不在精度或姿态判据（这两类合计只有 8 题），而在**走一半被冻住**：76% 的失败离目标还 >100 mm，其中多数是安全层反复全拒。这正是 `backoff`（全拒时缩步顶上）对着的那块；也说明 #15（候选数与安全档位）比"往打分里加剩余时限"更值钱。
|
||||
- **反直觉但自洽的一组数**：两段推理**开环更贴老师**（E1-b：0.43 vs 4.33）、**行程比更接近 1**（1.571 vs 2.445）、**到了的题还更快**（1.67 vs 2.31 s），但闭环到达掉 14.8 个百分点。逐题对照：108 题两者都到、**39 题只有 1 步能到、只有 1 题是 2 步独有** ⇒ 不是"换了一批题"，是同一批题里 1 步原本能做对的那部分被卡住了。
|
||||
- **卡点在安全层不在策略**：刹车从 19.25 → 86.3/题（4.5 倍），未到达组的刹车中位数从 45 → 176（几乎每步都被拒 ⇒ 原地冻结，末态误差 p50 也从 230 → 406 mm）。也就是说**发布的那 57.4% 是与"一步那套偏置且被削弱过的动作"共同标定出来的**：策略、反馈增益、认证口径、DAgger 采集分布四者都围着它。
|
||||
- **两种解释必须分开，否则下一步会走错**：
|
||||
|
||||
Reference in New Issue
Block a user