diff --git a/docs/index.md b/docs/index.md
index 6d9eacb..4952221 100644
--- a/docs/index.md
+++ b/docs/index.md
@@ -59,6 +59,14 @@ confirm the two sides talk to each other, not to train anything.
## A real VLA, end to end
+
+
+*The **unmodified** pi0.5, driven through the boundary on `libero_spatial`
+task 0 - three episodes, all three successful. This is the observation stream
+the policy itself sees, at its native 224x224, not an outside camera. The
+fine-tuned policy is the one that scores zero, further down.*
+
The same two processes carry a full-size pi0.5. The env client steps LIBERO,
the server answers with actions, and FPO trains on the feedback that comes
back. Nothing about the boundary changes; only the policy does.
diff --git a/docs/index.zh.md b/docs/index.zh.md
index e3e84c3..3c7e61e 100644
--- a/docs/index.zh.md
+++ b/docs/index.zh.md
@@ -53,6 +53,13 @@ dummy 算法的 `learn` 是一个 sleep，不会移动任何权重。它用来
## 真实 VLA，端到端
+
+
+*画面是**未经微调**的 pi0.5 通过这条边界在 `libero_spatial` 任务 0 上的执行过程——
+三个回合，全部成功。这是策略自己看到的观测流，原生 224×224，不是外部机位。
+微调之后成功率归零的那个策略，在下文。*
+
同样的两个进程也能承载一个完整尺寸的 pi0.5：环境端步进 LIBERO，服务端返回动作，
FPO 用回传的反馈训练。边界本身没有任何改动，变的只是策略。
diff --git a/docs/media/libero-base.mp4 b/docs/media/libero-base.mp4
new file mode 100644
index 0000000..8d29512
Binary files /dev/null and b/docs/media/libero-base.mp4 differ
diff --git a/docs/media/libero-demo.gif b/docs/media/libero-demo.gif
new file mode 100644
index 0000000..9df3509
Binary files /dev/null and b/docs/media/libero-demo.gif differ
diff --git a/docs/media/libero-wrist.mp4 b/docs/media/libero-wrist.mp4
new file mode 100644
index 0000000..95bf026
Binary files /dev/null and b/docs/media/libero-wrist.mp4 differ