PlugRL · tactino · Sep 27, 2026 · Sep 27, 2026 · Sep 27, 2026 · Sep 27, 2026
diff --git a/docs/index.md b/docs/index.md
@@ -100,8 +100,9 @@ episodes, and the clip is the episode with the median return, not the best
 one. `dppo-policy · FPO` is missing because it cannot exist: FPO trains flow
 policies only. The three pi0.5 clips all start from the same scene, the first
 one the released policy solves, and the numbers under them come from
-fifty-episode evaluations. pi0.5 falling to zero after one FPO iteration is a
-defect of ours we are still tracking down, not a finding about FPO. The
+fifty-episode evaluations. pi0.5 used to fall to zero after one FPO
+iteration; the defect was ours, in how our FPO scored an action chunk, and
+with FPO++'s way of scoring it the policy survives ([E32](https://github.com/PlugRL/plugrl-server/tree/main/experiments/e32-pi0-fpo-plus-plus)). The
 scripts that made all of this are in
 [figures/coverage](https://github.com/PlugRL/plugrl-server/tree/main/figures/coverage).
 
@@ -139,6 +140,13 @@ were pre-registered, and one of them is falsified. The numbers, the recorded
 environment of both processes, and what none of it supports:
 [E11](https://github.com/PlugRL/plugrl-server/tree/main/experiments/e11-vla-rl-libero).
 
+The collapse turned out to be ours. This FPO scored an action chunk by
+averaging the error over all 320 of its elements, most of them padding or
+steps the client never executed; scored as FPO++ scores it - the executed
+steps and the dimensions LIBERO uses - one update leaves pi0.5 at 33 of 50
+([E32](https://github.com/PlugRL/plugrl-server/tree/main/experiments/e32-pi0-fpo-plus-plus)). That it survives is all this shows: nothing yet says RL
+makes it better.
+
 ## The environment side needs neither CUDA nor a GPU
 
 The training server is 6.5G and wants a GPU. The machine running environments

diff --git a/docs/index.zh.md b/docs/index.zh.md
@@ -85,8 +85,9 @@ PlugRL 把策略、算法和环境拆开，所以真正值得回答的问题是
 每段视频都取自三个种子里最后十轮回报居中的那个种子的最终检查点。这个检查点评估了五
 个回合，放出来的是回报居中的那一回合，不是最好的那一回合。表里没有 `dppo-policy · FPO`，
 因为这个组合不存在：FPO 只能训练流策略。pi0.5 的三段视频都从同一个场景开始，也就是原
-版策略能完成的第一个场景；视频下面的数字来自 50 个回合的评估。pi0.5 一轮 FPO 后掉到
-零，是我们平台的问题，还在排查，不是关于 FPO 的结论。生成这些内容的脚本在
+版策略能完成的第一个场景；视频下面的数字来自 50 个回合的评估。pi0.5 以前一轮 FPO 后
+就掉到零，问题出在我们这边——这里的 FPO 给动作块打分的方式不对，换成 FPO++ 的打分方式后
+策略就不再被毁（[E32](https://github.com/PlugRL/plugrl-server/tree/main/experiments/e32-pi0-fpo-plus-plus)）。生成这些内容的脚本在
 [figures/coverage](https://github.com/PlugRL/plugrl-server/tree/main/figures/coverage)。
 
 训练端和环境端连 Python 环境都不共用。训练这些格子的服务端，没有一台装了 MuJoCo、robosuite
@@ -116,6 +117,11 @@ FPO 用回传的反馈训练。边界本身没有任何改动，变的只是策
 进程各自被记录下来的环境，以及这些数据**不能**支持的结论，见
 [E11](https://github.com/PlugRL/plugrl-server/tree/main/experiments/e11-vla-rl-libero)。
 
+这次归零后来查清了，原因在我们：这里的 FPO 给一个动作块打分时，对全部 320 个元素的
+误差取平均，其中大部分是补零的维度或客户端根本没执行的步。按 FPO++ 的方式打分——只算
+执行了的步、只算 LIBERO 用到的维度——一次更新后 pi0.5 还有 33/50（[E32](https://github.com/PlugRL/plugrl-server/tree/main/experiments/e32-pi0-fpo-plus-plus)）。
+这只说明它不再被毁，还不能说明强化学习能让它变好。
+
 ## 环境端既不需要 CUDA，也不需要 GPU
 
 训练端有 6.5G，而且要一张显卡；跑环境的那台机器不必如此。LIBERO 环境端可以装成

diff --git a/docs/javascripts/coverage.js b/docs/javascripts/coverage.js
@@ -49,7 +49,7 @@
       clientNote: (e) => `uv sync --extra ${e.extra}: ${e.packages}`,
       cmdHint:
         "Click another cell: the highlighted words are what changed. Shown with seed 0 " +
-        "(the experiments ran seeds 0-2); the client's episode count is only an upper bound, " +
+        "(each experiment ran three seeds); the client's episode count is only an upper bound, " +
         "the server ends the run.",
       vlaHint: "Cluster paths are placeholders; the linked script has the rest.",
       copy: "Copy",
@@ -87,7 +87,7 @@
       client: "环境客户端",
       clientNote: (e) => `uv sync --extra ${e.extra}：${e.packages}`,
       cmdHint:
-        "点另一格：高亮的就是变了的词。这里统一写成种子 0（实验跑的是 0 到 2 三个种子）；" +
+        "点另一格：高亮的就是变了的词。这里统一写成种子 0（每个实验跑了三个种子）；" +
         "客户端的回合数只是个上限，什么时候停由服务端决定。",
       vlaHint: "集群上的路径用占位符代替，完整内容见链接的脚本。",
       copy: "复制",

diff --git a/docs/media/coverage-grid.jpg b/docs/media/coverage-grid.jpg
diff --git a/docs/media/coverage/coverage.json b/docs/media/coverage/coverage.json
diff --git a/docs/media/coverage/dppo-dppo-hopper.jpg b/docs/media/coverage/dppo-dppo-hopper.jpg
diff --git a/docs/media/coverage/dppo-dppo-hopper.mp4 b/docs/media/coverage/dppo-dppo-hopper.mp4
diff --git a/docs/media/coverage/dppo-dppo-square.jpg b/docs/media/coverage/dppo-dppo-square.jpg
diff --git a/docs/media/coverage/dppo-dppo-square.mp4 b/docs/media/coverage/dppo-dppo-square.mp4
diff --git a/docs/media/coverage/dppo-dppo-walker.jpg b/docs/media/coverage/dppo-dppo-walker.jpg
diff --git a/docs/media/coverage/dppo-dppo-walker.mp4 b/docs/media/coverage/dppo-dppo-walker.mp4
diff --git a/docs/media/coverage/fpo-dppo-cheetah.jpg b/docs/media/coverage/fpo-dppo-cheetah.jpg
diff --git a/docs/media/coverage/fpo-dppo-cheetah.mp4 b/docs/media/coverage/fpo-dppo-cheetah.mp4
diff --git a/docs/media/coverage/fpo-dppo-hopper.jpg b/docs/media/coverage/fpo-dppo-hopper.jpg
diff --git a/docs/media/coverage/fpo-dppo-hopper.mp4 b/docs/media/coverage/fpo-dppo-hopper.mp4
diff --git a/docs/media/coverage/fpo-dppo-walker.jpg b/docs/media/coverage/fpo-dppo-walker.jpg
diff --git a/docs/media/coverage/fpo-dppo-walker.mp4 b/docs/media/coverage/fpo-dppo-walker.mp4
diff --git a/docs/media/coverage/pi0-fpo.jpg b/docs/media/coverage/pi0-fpo.jpg
diff --git a/docs/media/coverage/pi0-fpo.mp4 b/docs/media/coverage/pi0-fpo.mp4