diff --git a/agentlightning/verl/rollout_adapter.py b/agentlightning/verl/rollout_adapter.py
index 18f2e1cca..a8dda25fe 100644
--- a/agentlightning/verl/rollout_adapter.py
+++ b/agentlightning/verl/rollout_adapter.py
@@ -9,6 +9,7 @@
 import json
 import zipfile
 from typing import Any, cast
+from urllib.request import url2pathname
 
 import numpy as np
 import torch
@@ -247,7 +248,7 @@ def _load_pil_image(url: str) -> Any:
         _, _, payload = url.partition(",")
         return Image.open(io.BytesIO(base64.b64decode(payload))).convert("RGB")
     if url.startswith("file://"):
-        return Image.open(url[len("file://") :]).convert("RGB")
+        return Image.open(url2pathname(url.removeprefix("file:"))).convert("RGB")
     if url.startswith(("http://", "https://")):
         # [multimodal-patch] Remote images are fetched from the trainer process with
         # content-type and size validation.
diff --git a/docs/80-example-multimodal-qa.md b/docs/80-example-multimodal-qa.md
index 6cd6e9e5c..eb6ecb970 100644
--- a/docs/80-example-multimodal-qa.md
+++ b/docs/80-example-multimodal-qa.md
@@ -16,6 +16,8 @@ The dataset is generated with PIL at startup — there is nothing to download:
 - The agent sends the image (as a base64 `image_url` content part) plus the question "How many red circles are in the image?" to the proxied VLM.
 - A rule-based reward scores the answer: 1.0 if the first integer in the reply equals the true circle count, 0.0 otherwise.
 
+For images already stored on the trainer host, construct each `image_url` with `Path("images/frame 1.png").resolve().as_uri()` (from `pathlib`). Local file URIs preserve spaces, Unicode, and literal percent characters in filenames.
+
 Any mrope VLM supported by verl works; pass `--model` to change the default model.
 
 ## Training
diff --git a/tests/verl/test_rollout_adapter.py b/tests/verl/test_rollout_adapter.py
index 8d0843926..dc289ae9a 100644
--- a/tests/verl/test_rollout_adapter.py
+++ b/tests/verl/test_rollout_adapter.py
@@ -7,6 +7,7 @@
 import json
 import sys
 import zipfile
+from pathlib import Path
 from types import SimpleNamespace
 from typing import ClassVar
 
@@ -594,6 +595,38 @@ def test_load_pil_image_decodes_data_url() -> None:
     assert image.mode == "RGB"
 
 
+@pytest.mark.parametrize("filename", ["plain.png", "frame one.png", "图像.png", "frame%20.png"])
+def test_load_pil_image_decodes_file_uri(tmp_path: Path, filename: str) -> None:
+    from PIL import Image
+
+    image_path = tmp_path / filename
+    Image.new("RGB", (2, 2), (255, 0, 0)).save(image_path)
+    # A second unquote would load the wrong image for the literal-percent case.
+    if filename == "frame%20.png":
+        Image.new("RGB", (2, 2), (0, 0, 255)).save(tmp_path / "frame .png")
+
+    image = rollout_adapter_module._load_pil_image(image_path.as_uri())
+
+    assert image.size == (2, 2)
+    assert image.mode == "RGB"
+    assert image.getpixel((0, 0)) == (255, 0, 0)
+
+
+def test_file_uri_image_row_is_kept_for_training(tmp_path: Path) -> None:
+    from PIL import Image
+
+    image_path = tmp_path / "frame 50%.png"
+    Image.new("RGB", (2, 2), (255, 0, 0)).save(image_path)
+    rollout = _transition_rollout([_image_triplet([1, 2], [3], image_urls=[image_path.as_uri()])])
+
+    batch, _ = _transition_adapter(FakeMropeProcessor()).get_train_data_batch([rollout])
+
+    assert batch.batch["is_drop_mask"].tolist() == [False]
+    assert batch.non_tensor_batch["multi_modal_inputs"][0] is not None
+    keep = (~batch.batch["is_drop_mask"].bool()).nonzero(as_tuple=True)[0].tolist()
+    assert batch[keep].non_tensor_batch["rollout_id_list"].tolist() == [rollout.rollout_id]
+
+
 def test_load_pil_image_fetches_remote_url_by_default(monkeypatch: pytest.MonkeyPatch) -> None:
     pytest.importorskip("PIL")
     httpx = pytest.importorskip("httpx")