Build reports

后伸悬臂 46→56mm，后壳跟着后移。七件三视图 + 全部 STL 下载， 衔接面三种接法讲清楚（借用相机原孔 / 热熔螺母 / 长圆槽）， 每颗螺丝按 1mm=6px 等比画出来，含热熔螺母规格和烫装手法。 六项干涉全空，力矩 0.245→0.320 kgf·cm（舵机余量 ~47×）。

iPhone 13 减压壳。三层夹持 + 统一蘑菇头接口：底板是通用插座（永远不用重打）， 包子是可换零件，换一套布局只重打一张面板。尺寸全部取自 Apple 官方工程图。

82A 料找回来了 —— 就是设计基准料，所以三个长度（89.4 / 100 / 120）全部回到基线几何， 零补偿。肋距改成 3.6：在生成体上实测 4.37mm，对上原厂红爪卡尺实测的 4.337mm。 摩擦花纹在成品 STL 上验证满铺（尖端段面片密度 7 倍于无花纹本体）。 含按 82A 调整过的切片参数和打完要做的三项测量。

两个可打印件，把「捏一下感觉太软」变成 N/mm。悬臂 + 挂砝码（不是用手拉弹簧秤： 手会抖、方向不定、读数还在动）。装夹方式和真机完全一致，所以量到的就是装在机器人上的刚度。 含刚度 / 回弹 / 摩擦三套测法、选重量的判据、以及一条必读的材料警告。

起因是一个约束：这台机器人只有两个爪，加不了第三个。 这一页把项目里已经自己量出来的六条设计原理整理成册， 点出现在唯一缺的那件东西（一个几十块钱的测量装置）， 列出必须真正理解的四个核心概念、值得研究的实验室与设计、书和视频， 以及一条「用这个项目本身当训练场」的路线。

起因是一个约束：这台机器人只有两个爪，加不了第三个。这一页是那个问题的答案 —— 本项目里自己量出来的六条设计原理、必须真正理解的四个核心概念（力/形封闭、欠驱动、 顺应机构、Fin-Ray 效应）、值得看的实验室与设计、以及现在唯一缺的那件东西：一个测量装置。

把训练好的 checkpoint 放到真手臂上跑 10 次，整个循环在手机面板上按按钮完成，不用终端。 为什么做成面板（理由是安全不是方便）、安全状态那四个词分别意味着什么该怎么反应、 从三条自检到判定结果的一步步流程，以及做这个时抓到的两个 bug。 站在机器人旁边时打开的就是这一页。

Self-contained write-up in English for an outside reviewer. What was built and why, calipered geometry with a self-check, an unresolved URDF/measurement gap, rib parameters and the checker that had never run, the friction pattern and one CGAL trap, the Shore-A material table against hardware feedback, and the snap-on leaf spring. Two open questions at the end.

第三版。修正了一个理解错误：刀片是绕着杯子的圆周包过去的，杯子轴线垂直于刀片平面。 所以几字板簧的自由腿要沿杯周继续包 —— 单片补上 80°，两片合计把包覆从 57% 拉到约 100%。 末端钩的两个方向都出了 STL。

What may scale with blade length and what is pinned to the hardware, the one quantity that must be back-solved or the mount fit breaks, and the twelve checks to run after changing it.

输入想要的长度（或杯子直径）和打印材料，自动完成参数映射 → STL → 网格检查 → 尺寸复核 → 三视图渲染 → 工程图 → HTML 报告。材料硬度到结构参数的换算、 可打印性下限、每一项必须复核的尺寸，全部固化在代码里。

九个候选件全部 check_mesh PASS。柔性总账、加长的好处与代价、推荐件三视图与工程图、 以及摩擦花纹从尖锥改成平顶台地的依据（真实接触面积 ≈0 → 61%）。

第一对打印件几乎无法弯折、容易打滑。把材料和结构两个嫌疑分开量：新件材料反而多 12%、 内梁更薄，几何上该更软 —— 锅在 TPU 98A（设计基准是 95A，刚度约 2.6 倍）。95A 拿不到， 附用结构调软的方案，按改动量从小到大排。

Rebuilt: the tip draws in on a parabola, the ribs reach 10 mm further down, and length is now a free parameter that keeps the mount fit intact. Drawing, renders, print settings, and STLs at 89.4 / 100 / 120 mm to download.

Five caliper readings came back. Two self-checks passed, one was revised — and the result overturns what the field-test document blamed: the crossing fingers were never a hole problem. With annotated drawings of exactly which points these are.

The standing workflow for every future component — what to measure, the confirmation gate that has to pass before any CAD is written, and how to tell a real constraint from a guess. Each stage named after the failure that justifies it.

What two-segment and three-segment actually look like — four profiles drawn to scale plus renders, now that the protrusion feature is measured at 10–12 mm.

Six named points, five caliper readings, all in one closed state — enough to settle the hole shift, the convergence angle, and whether the mating face needs flattening at all.

The long-jaw Fin-Ray gripper blade, rebuilt from field-test feedback. Seven rendered views off the shipping mesh, what changed from v3, and the two dimensions still waiting on a caliper.

Top view of the gripper closing, simulated from the robot's own URDF joint definition. Drag to sweep 0° to 97.4°.

The 68 mm file the whole v4 chain is built on, that nobody remembers supplying — rendered, measured, and traced through every claim the repo makes about it.

Why the v4 tip protrusion is a trapezoid instead of a continuous taper, and three continuous alternatives rendered from real meshes — including the barb whose hook face sits square to the inner rail.

Training and simulation need the dataset, not the robot — so they can run on anyone's GPU. How to rent one step by step, what it actually costs, and the Windows build worth buying only if renting stops making sense.

ACT vs SmolVLA on the identical dataset — why their loss numbers cannot be compared, what went wrong in training versus what was just a full disk, and how to tell next time whether a run needs redoing. Includes a plain-language glossary and both papers.

The cup-grasp policy's training curve, and why a final loss of 0.138 is not evidence the policy learned the task. Interactive — drag across any chart to read values.