Solo · AI-assisted (Claude Code) · In progress独立 · AI 协助开发（Claude Code）· 开发中
VR Sheepdog
Unity 6 · C# · Flocking AI · A* · Deterministic Sim
You are the shepherd, not the dog. Seated in your own living room, you give hand signals to an autonomous sheepdog that herds a flock around your real sofa and table into a pen. The simulation core is done; the Quest mixed-reality shell is next. The commands are named after real handler calls: Walk Up, Lie Down, Come Bye, Away and That'll Do. 你是牧羊人，不是狗。坐在自己的客厅里，用手势指挥一条自主牧羊犬，把羊群绕过真实的沙发和茶几赶进围栏。模拟核心已经完成，下一步是 Quest 混合现实外壳。指令沿用真实牧羊比赛的口令：Walk Up、Lie Down、Come Bye、Away、That'll Do。
Problem问题
In a real room, furniture blocks the flock and the dog. A dog that works on an open field can fail completely around a sofa.
在真实房间里，家具会挡住羊群和狗。在空地上能干活的狗，到了沙发旁边可能完全失灵。
Decision决策
Build a deterministic, seedable simulation first, test it over ten seeds per layout, and fix each failure at its root with a test.
先做一个确定性、可指定种子的模拟，每种场景跑十个种子，每个失败都找到根因并补上测试。
Outcome结果
10 of 10 seeds pen all 12 sheep on both the open field and the L-shaped sofa room; 44 EditMode tests pass.
空地和 L 形沙发房间两种场景，10 个种子都把 12 只羊全部赶进围栏；44 个 EditMode 测试通过。
Top-down debug render of the L-shaped sofa room (seed 5). Black: dog. White: sheep. The line is the dog's planned path.L 形沙发房间的俯视调试画面（种子 5）。黑色是狗，白色是羊，线条是狗规划的路线。
How the herd moves羊群怎么动
Sheep羊
A flocking model from herding research (Strömbom et al. 2014): cohesion with six neighbours, separation, and fleeing pressure that grows as the dog gets closer.
采用放牧研究中的群体模型（Strömbom 等，2014）：和六个邻居聚拢、保持间距，狗越近逃离的压力越大。
Dog狗
Collects strays, then drives the group toward the pen, unless the player's command overrides it.
先收拢掉队的羊，再把羊群往围栏赶；玩家下指令时以指令为准。
Room房间
A 5 cm occupancy grid that knows each body's size, so nothing squeezes through a gap narrower than itself; bodies slide along walls instead of sticking.
5 厘米一格的占用网格，按身体大小判断能否通过，比身体窄的缝钻不过去；碰到墙会沿墙滑动而不是卡住。
The simulation steps at a fixed 30 Hz and is deterministic for a seed and a command log. A five-second snapshot buffer already supports rewind and fork: go back, give a different command, compare.模拟以固定 30 Hz 步进，给定种子和指令记录结果完全确定。五秒快照缓冲已经支持倒带和分叉：退回去，换一个指令，对比结果。
From 1 in 10 to 10 in 10 around a sofa绕沙发：从 10 次成功 1 次到 10 次全成
The L-shaped room started at 1 success in 10 seeds. Five separate root causes were behind it, and each got a test: pathfinding returned nothing when it started inside an inflated obstacle; the flock's path used twice the sheep radius, so a 0.345 m gap was impassable; the drive point was purely geometric and landed behind the sofa arm; the pen entry point was computed behind the sofa; and "gathered" only checked distance, so a flock split by the sofa arm counted as one group.
L 形房间一开始 10 个种子只成功 1 次。背后有五个独立的根因，每个都补了测试：起点落在障碍膨胀区内时寻路直接返回空；羊群路径用了两倍羊身半径，0.345 米的缝变得过不去；驱赶站位只按几何算，落到了沙发扶手后面；围栏入口点算在了沙发后面；「已聚拢」只看距离，被沙发扶手隔开的两群羊也被当成一群。
The dog's outrun was then planned with A* over a ground cost: cells next to sheep and cells in front of the flock cost more, so the dog goes round behind the flock instead of cutting through it. I first tried the wide arc used in real trials; on the open field it fell to 7 of 10, because in a small room the sheep feel the dog everywhere. A tight, fast pass behind the flock won.
之后狗的绕跑改用 A* 按「地面代价」规划：贴着羊的格子和羊群前方的格子代价更高，狗会从羊群后面绕过去，而不是从中间穿过。我先试了真实比赛里的大弧线绕跑，结果空地成功率掉到 7/10：小房间里羊在哪儿都能感到狗。最后定为紧贴羊群后方快速绕过。
- Open field: 10 of 10 seeds penned all 12 sheep, 15.8–32.1 s, mean 18.6 s.
- 空地：10 个种子都把 12 只羊全部赶进围栏，用时 15.8–32.1 秒，平均 18.6 秒。
- L-shaped sofa room: 10 of 10, 27.4–43.1 s, mean 35.2 s.
- L 形沙发房间：10 个种子全部成功，用时 27.4–43.1 秒，平均 35.2 秒。
Ten seeds per layout in the Unity EditMode run.Unity EditMode 中每种场景跑 10 个种子。
Tooling and what's next工具与下一步
The core is plain C# with no Unity dependency, so a .NET 9 harness compiles it outside the editor to run ten-seed tables, per-second traces and pathfinding probes in seconds. An editor recorder renders any seed to video. Next: a side-by-side rewind and fork view (M2), then the Quest shell, where the scanned room becomes the occupancy grid and three hand gestures replace the keyboard (M3).
核心是不依赖 Unity 的纯 C#，所以 .NET 9 的离线工具可以在编辑器外编译它，几秒内跑完十个种子的结果表、逐秒轨迹和寻路探针。编辑器录制器可以把任意种子渲染成视频。下一步：倒带与分叉的并排对比视图（M2），然后是 Quest 外壳：扫描到的房间变成占用网格，三个手势代替键盘（M3）。
Open field, seed 3, fully autonomous.空地，种子 3，全自动。