PlugRL · tactino · Sep 29, 2026 · Sep 29, 2026
diff --git a/docs/media/coverage/coverage.json b/docs/media/coverage/coverage.json
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
diff --git a/docs/vla.md b/docs/vla.md
@@ -3,14 +3,17 @@
 The home page shows a full-size pi0.5 running across PlugRL's boundary, with
 its evaluations matching openpi's. This page is the rest of that record: what
 fine-tuning it with reinforcement learning through PlugRL has done so far.
-In short, it has not made the policy better.
+In short, it has not made the policy better. The collapses the first runs
+showed were defects of ours, and with them fixed FPO leaves the policy about
+where it started.
 
 <div class="cov" data-part="vla" data-src="/media/coverage/coverage.json"></div>
 
 All three clips start from the same scene, the first one the released policy
 solves, and the numbers under them come from fifty-episode evaluations. Over
-seven such evaluations the released policy scored between 28 and 37. Click a
-clip to see the two commands behind it.
+seven such evaluations the released policy scored between 28 and 37. Where an
+experiment ran two seeds, the clip and the number are the lower-scoring
+seed's. Click a clip to see the two commands behind it.
 
 ## How it went
 
@@ -32,8 +35,14 @@ than FPO++'s. Ten DPPO iterations left it at 20, three standard deviations
 below the released policy's mean
 ([E36](https://github.com/PlugRL/plugrl-server/tree/main/experiments/e36-pi0-longer)).
 
-A run with FPO++'s fine-tuning in full is under way, to check whether what
-remains of the collapse is still ours.
+That fall was ours too. With the rest of FPO++'s fine-tuning in place (its
+optimizer, gradient clipping, critic learning rate, raw rewards, lambda and
+clip), ten FPO iterations left the same two seeds at 40 and 27 of 50
+([E42](https://github.com/PlugRL/plugrl-server/tree/main/experiments/e42-pi0-fpo-plus-plus)).
+Neither reaches the bar set before the run, 42 of 50, so this is FPO holding
+pi0.5, not improving it. The lower seed, whose clip is above, was drifting
+the way E36's runs fell, only far more slowly. Whether it holds past ten
+iterations is open.
 
 The scripts that recorded the clips are in
 [figures/coverage](https://github.com/PlugRL/plugrl-server/tree/main/figures/coverage).
diff --git a/docs/vla.zh.md b/docs/vla.zh.md
@@ -2,13 +2,14 @@
 
 主页展示了一个完整尺寸的 pi0.5 穿过 PlugRL 的边界跑通，评估结果和 openpi 公布的一致。
 这一页是那份记录的其余部分：通过 PlugRL 用强化学习微调它，到目前为止做到了什么。
-简单说，还没能让策略变好。
+简单说，还没能让策略变好。前几次跑出来的崩塌都是我们自己的缺陷，修掉之后 FPO 基本让策略
+停在原来的水平。
 
 <div class="cov" data-part="vla" data-src="/media/coverage/coverage.json"></div>
 
 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。原版策略在七次这样的评估里得分在 28 到 37 之间。点一下视频，会显示它背后的
-两条命令。
+回合的评估。原版策略在七次这样的评估里得分在 28 到 37 之间。实验跑了两个种子的，视频和
+数字都取分数较低的那个。点一下视频，会显示它背后的两条命令。
 
 ## 经过
 
@@ -26,7 +27,11 @@
 还是自己的默认设置，没换成 FPO++ 的；十轮 DPPO 后是 20/50，比原版策略的平均低了三个
 标准差（[E36](https://github.com/PlugRL/plugrl-server/tree/main/experiments/e36-pi0-longer)）。
 
-换上 FPO++ 完整设置的实验正在跑，用来确认剩下的崩塌还是不是我们的问题。
+这次下跌也是我们的问题。把 FPO++ 其余的微调设置补齐（优化器、梯度裁剪、critic 学习率、
+原始奖励、lambda 和 clip）之后，同样两个种子跑十轮 FPO，分别是 40/50 和 27/50
+（[E42](https://github.com/PlugRL/plugrl-server/tree/main/experiments/e42-pi0-fpo-plus-plus)）。
+两个都没达到跑之前定好的 42/50，所以这是 FPO 稳住了 pi0.5，不是把它变好了。分数较低的
+那个种子（上面的视频就是它）在往 E36 那样的方向漂，只是慢得多；再往后跑还稳不稳，还不知道。
 
 录这些视频的脚本在
 [figures/coverage](https://github.com/PlugRL/plugrl-server/tree/main/figures/coverage)。