训练好的 checkpoint 第一次上真手臂。 9 个批次、31 次计入试验、0 次成功 —— 但数据本身讲了一个很清楚的故事。
| 安全层判定 | 帧数 | 占比 |
|---|---|---|
| OK 原样发出 | 5468 | 78.9% |
| CLIPPED 被钳位 | 1455 | 21.0% |
| HELD 被拒绝 | 3 | 0.0% |
这条排除了一整类原因。总共 6926 帧，安全层一帧都没拒 —— 策略输出的动作全部在笛卡尔包络内、没有 NaN、没有绝荒量级、没有超限位。 也就是说失败不是因为安全层拦着，也不是策略在乱输出。 中止次数同样是 0，没有任何一次因为「连续 15 帧被拒」而提前结束。
| speed | 试验次数 | 钳位率 |
|---|---|---|
| 0.3 | 12 | 47.7% |
| 0.6 | 3 | 7.5% |
| 1.0 | 16 | 2.7% |
speed 0.3 时 47.7% 的帧被钳位，speed 1.0 时只有 2.7%。
原因在 run_policy_trials.py： --speed 是直接乘在步长上限上的 （filt.step_cap[k] *= a.speed）。 0.3 档把每帧允许的关节位移压到三分之一，于是策略想走的大部分步子都被削平了 —— 手臂实际执行的轨迹和策略想要的轨迹差很远，这本身就会造成失败。
推论：低速档并不是「更安全的同一个动作」，而是另一个被扭曲过的动作。 要判断策略本身行不行，应该看 speed 1.0 那 16 次 —— 那时钳位率只有 2.7%， 手臂执行的基本就是策略的原始意图。而它仍然 0/16。
| 批次时间戳 | speed | 试验数 | 成功 | 不计 | 钳位率 |
|---|---|---|---|---|---|
| 20260823-021014 | 0.3 | 1 | 0 | 1 | 64.2% |
| 20260823-022633 | 0.3 | 1 | 0 | 0 | 55.4% |
| 20260823-023122 | 0.3 | 1 | 0 | 0 | 49.3% |
| 20260823-024937 | 0.3 | 9 | 0 | 0 | 45.8% |
| 20260823-025232 | 0.6 | 1 | 0 | 0 | 7.1% |
| 20260823-025528 | 0.6 | 2 | 0 | 1 | 7.7% |
| 20260823-030334-recovered | 1.0 | 5 | 0 | 1 | 5.3% |
| 20260823-034134 | 1.0 | 10 | 0 | 0 | 0.1% |
| 20260823-044302 | 1.0 | 1 | 0 | 1 | 0.0% |
数据源：05-training/trials/trials-*.json， 由 run_policy_trials.py 每批次自动写出。
已排除：
仍待验证（按可能性排）：
整个 Space 的全部 76 次更新，最新在上。输入关键词可过滤。
| 时间 | 更新内容 |
|---|---|
| 2026-08-23 01:45 | replay verdict: geometry unchanged; gripper friction is the new factor |
| 2026-08-23 01:21 | add: standard practice in the literature vs our situation |
| 2026-08-23 01:15 | trials day 1: data-coverage evidence (r=-0.88), full tuning workflow |
| 2026-08-23 01:08 | trials day 1: 0/14 failure analysis, grasp sight-picture sheet, replay experiment |
| 2026-08-23 00:16 | print pack: complete parameter table for PLA+ standard / PLA+ fast / clear PLA |
| 2026-08-22 23:08 | print pack: add speed settings section - what to raise, what not to touch |
| 2026-08-22 22:31 | cat ears: bezier profile with real curvature, roots sunk into the base |
| 2026-08-22 22:20 | homepage: live panel links (no password) |
| 2026-08-22 22:17 | head v12 print pack: all 10 STLs, three-views, full print & assembly guide |
| 2026-08-22 22:10 | ear pod: speaker corner bosses at measured 41.5mm spacing (was 43 by eye) |
| 2026-08-22 22:01 | head v12: fix dovetail slots never cut, pod insert holes drilled outward, slim ear pod 28% |
| 2026-08-22 21:30 | head v12: fix blocked side vent (annular pad), face slot 20x10, cat ears + socket covers |
| 2026-08-22 21:06 | head v12: three-board mounting (BT vertical + HU-004 flat), core_link untouched |
| 2026-08-22 20:38 | head v12: speaker back-vent aligned, M3 board posts, top dovetail sockets + key |
| 2026-08-22 20:01 | head v12: full assembly page - three-views, all 7 STLs, joint & screw guide |
| 2026-08-22 19:42 | popcase solo: deepen cavity by the standoff, add camera rim; rebuilt STL |
| 2026-08-22 19:27 | 82A pack: recommend 100mm with the wrap-vs-grip-force reasoning; fix the OEM-length error |
| 2026-08-22 19:25 | restore latest popcase page (20:58, from the other session) + its STLs; space dir is now the single source |
| 2026-08-22 19:21 | head beam +10mm page; popcase card moved up front |
| 2026-08-22 19:10 | 82A pack: L120 rib check completed, all pass |
| 2026-08-22 19:10 | 82A print pack: three lengths at OEM rib pitch, verified texture coverage |
| 2026-08-22 19:03 | trials panel: add the actual URLs to open |
| 2026-08-22 18:52 | add blade test rig: mount + weight hanger, and the measurement method |
| 2026-08-22 18:35 | add gripper design learning path |
| 2026-08-22 18:23 | add trials panel operating guide |
| 2026-08-22 17:58 | Blade v4 spec sheet and ACT convergence report |
| 2026-08-22 17:20 | Blade v4 spec sheet and ACT convergence report |
| 2026-08-22 17:18 | review page: drop internal process detail per operator |
| 2026-08-22 16:42 | review page: English, lead with the work, questions last |
| 2026-08-22 16:40 | gripper review page: rewrite in English for the outside reviewer |
| 2026-08-22 16:38 | Blade v4 spec sheet and ACT convergence report |
| 2026-08-22 16:33 | tip clip: wave corrugation + transverse friction ridges |
| 2026-08-22 16:29 | add standalone gripper review collection for external reviewer |
| 2026-08-22 16:09 | tip clip: final printable set (rail-following legs, hook clearance check) |
| 2026-08-22 15:57 | tip clip: printable set, closed hat, reverted leg direction, thickness ladder |
| 2026-08-22 15:40 | tip clip v3: free leg follows the cup circumference; both hook directions |
| 2026-08-22 15:23 | tip clip v2: 几-shaped leaf spring, in the blade plane |
| 2026-08-22 15:13 | tip C-clip: first pass for operator verification |
| 2026-08-22 12:38 | Blade v4 spec sheet and ACT convergence report |
| 2026-08-22 09:34 | Blade v4 spec sheet and ACT convergence report |
| 2026-08-22 09:30 | Blade v4 spec sheet and ACT convergence report |
| 2026-08-22 08:39 | Blade v4 spec sheet and ACT convergence report |
| 2026-08-22 08:35 | head v1: flange capped at 12.5mm tall / 5mm inward per operator limits, wrap lips removed, all eight checks pass |
| 2026-08-22 08:05 | head v1: real interference checks (the /tmp include ones were false negatives), five geometry bugs fixed, 4-point ear mount, counterbored holes |
| 2026-08-22 07:44 | Blade v4 spec sheet and ACT convergence report |
| 2026-08-22 07:33 | head: add a dimensioned three-view of the existing camera from the operator measurements |
| 2026-08-22 07:31 | head v1: flange full-height with wrap lips, rotation load analysis, per-part dimensions and interface map |
| 2026-08-22 07:08 | Blade v4 spec sheet and ACT convergence report |
| 2026-08-22 06:41 | head v1: fix crossbar poking through the roof and the 12.6mm gap at the shell mount; head resized to 121x95x100; seven-way interference matrix |
| 2026-08-22 06:40 | Blade v4 spec sheet and ACT convergence report |
| 2026-08-22 00:28 | head v1: gimbal bridge with Z-profile driver access, cradle, cable clip, ear pods for the measured speakers; drawings, exploded, STLs, print guide; single entry on the index |
| 2026-08-21 23:42 | head V12: corrected interface dims, gimbal bridge, camera cradle, cable clip, cutaway; linked from index |
| 2026-08-21 23:24 | Blade v4 spec sheet and ACT convergence report |
| 2026-08-21 23:19 | head V12: add whole-head hero render |
| 2026-08-21 23:18 | head V12: dimensioned three-view drawings, screw-hole schedule, shared-screw stack-up |
| 2026-08-21 23:08 | head V12: core link part, three-views, exploded, STLs and print guide |
| 2026-08-21 22:47 | head: engineering redraw of the operator's 2026-08-22 sketch |
| 2026-08-21 22:00 | Blade v4 spec sheet and ACT convergence report |
| 2026-08-21 21:25 | Blade v4 spec sheet and ACT convergence report |
| 2026-08-21 20:58 | Blade v4 spec sheet and ACT convergence report |
| 2026-08-21 20:45 | Blade v4 spec sheet and ACT convergence report |
| 2026-08-21 20:35 | Blade v4 spec sheet and ACT convergence report |
| 2026-08-21 20:20 | Blade v4 spec sheet and ACT convergence report |
| 2026-08-21 20:14 | Blade v4 spec sheet and ACT convergence report |
| 2026-08-21 19:37 | Blade v4 spec sheet and ACT convergence report |
| 2026-08-21 19:11 | Blade v4 spec sheet and ACT convergence report |
| 2026-08-21 18:13 | Blade v4 spec sheet and ACT convergence report |
| 2026-08-21 17:51 | Blade v4 spec sheet and ACT convergence report |
| 2026-08-21 17:30 | Blade v4 spec sheet and ACT convergence report |
| 2026-08-21 17:18 | Blade v4 spec sheet and ACT convergence report |
| 2026-08-21 16:17 | Blade v4 spec sheet and ACT convergence report |
| 2026-08-21 16:04 | Blade v4 spec sheet and ACT convergence report |
| 2026-08-21 16:04 | Blade v4 spec sheet and ACT convergence report |
| 2026-08-21 15:52 | Blade v4 spec sheet and ACT convergence report |
| 2026-08-21 13:45 | Blade v4 spec sheet and ACT convergence report |
| 2026-08-21 13:45 | initial commit |