diff --git a/robots/libero/env_server.py b/robots/libero/env_server.py
index 2dd6aacb4..968c79a6e 100644
--- a/robots/libero/env_server.py
+++ b/robots/libero/env_server.py
@@ -101,6 +101,26 @@ def build_env_cfg(
     return cfg
 
 
+def _empty_init_state_tasks(suite) -> list[int]:
+    """Best-effort scan for tasks shipping an empty init-state set.
+
+    Only used to build a diagnostic message: a task whose init-state file
+    cannot be read is skipped here — it fails on its own once requested.
+    """
+    try:
+        num_tasks = suite.get_num_tasks()
+    except Exception:
+        return []
+    empty = []
+    for task_id in range(num_tasks):
+        try:
+            if len(suite.get_task_init_states(task_id)) == 0:
+                empty.append(task_id)
+        except Exception:
+            continue
+    return empty
+
+
 def make_env(
     task_id: int,
     seed: int,
@@ -114,6 +134,21 @@ def make_env(
     suite = _bench_mod.get_benchmark(suite_name)()
     first_id = sum(len(suite.get_task_init_states(t)) for t in range(task_id))
     trials = len(suite.get_task_init_states(task_id))
+    if trials == 0:
+        # Some shipped perturbation suites contain tasks whose pruned_init
+        # file holds zero states (RLinf/RPent#246). Such a task cannot be
+        # reset; continuing would die on a modulo by zero or — worse, in
+        # aggregation paths — silently shrink the evaluation denominator
+        # (e.g. 100 -> 90), making results incomparable. Fail loudly with
+        # every affected task so the run can be fixed up front.
+        empty = _empty_init_state_tasks(suite)
+        raise RuntimeError(
+            f"LIBERO suite '{suite_name}' task {task_id} has 0 init states "
+            "and cannot be reset or evaluated. Tasks with empty init-state "
+            f"sets in this suite: {empty if empty else 'unknown'}. Restore "
+            "the init files (see robots/libero/guides/pro_hybrid_guide.md "
+            "section 2.3) or exclude these tasks from the run."
+        )
     rid = first_id + (seed % trials)
     cfg = build_env_cfg(
         task_suite_name=suite_name,
diff --git a/robots/libero/guides/pro_hybrid_guide.md b/robots/libero/guides/pro_hybrid_guide.md
index 8a1791a81..914baf1fb 100644
--- a/robots/libero/guides/pro_hybrid_guide.md
+++ b/robots/libero/guides/pro_hybrid_guide.md
@@ -155,6 +155,19 @@ snapshot_download(repo_id='zhouxueyang/LIBERO-Pro', repo_type='dataset',
                   allow_patterns=['bddl_files/**','init_files/**'])"
 ```
 
+#### Known empty `pruned_init` files in the git repo (RLinf/RPent#246)
+
+These tasks' git-repo `pruned_init` files contain **0 states**, so they cannot
+be reset. The env server refuses to start on them (naming every affected task)
+instead of silently dropping them from the evaluation denominator. Re-syncing
+the HF snapshot above over the install regenerates these files:
+
+| Suite | Task id | Init file |
+|---|---|---|
+| `libero_10_task` | 2 | `KITCHEN_SCENE3_turn_on_the_stove_and_put_the_moka_pot_on_it.pruned_init` |
+| `libero_spatial_task` | 3 | `pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate.pruned_init` |
+| `libero_spatial_task` | 7 | `pick_up_the_black_bowl_on_the_stove_and_place_it_on_the_plate.pruned_init` |
+
 ### 2.4. Verify
 
 ```bash
diff --git a/tests/unit_tests/robots/libero/test_env_server_init_states.py b/tests/unit_tests/robots/libero/test_env_server_init_states.py
new file mode 100644
index 000000000..0163ee94f
--- /dev/null
+++ b/tests/unit_tests/robots/libero/test_env_server_init_states.py
@@ -0,0 +1,120 @@
+# Copyright 2026 The RPent Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Offline tests for ``make_env`` init-state handling (RLinf/RPent#246).
+
+Shipped perturbation suites can contain tasks whose ``pruned_init`` file
+holds zero states. ``make_env`` must refuse to build such an env with a
+diagnostic naming every affected task, instead of dying on a modulo by
+zero or silently shrinking the evaluation denominator.
+"""
+
+from __future__ import annotations
+
+import sys
+from types import ModuleType
+
+import pytest
+
+import robots.libero.env_server as env_server
+
+
+class FakeSuite:
+    """Mimics the LIBERO benchmark API surface ``make_env`` consumes."""
+
+    def __init__(self, counts: list[int]):
+        self._counts = counts
+
+    def get_num_tasks(self) -> int:
+        return len(self._counts)
+
+    def get_task_init_states(self, task_id: int) -> list[object]:
+        return [object()] * self._counts[task_id]
+
+
+def _install_fake_rlinf(monkeypatch: pytest.MonkeyPatch, suite: FakeSuite):
+    """Stub the ``rlinf`` benchmark/env modules ``make_env`` imports.
+
+    Returns ``(env_cls, captured_cfgs)`` so tests can inspect the env cfg
+    that a successful ``make_env`` hands to ``LiberoEnv``.
+    """
+    captured_cfgs: list = []
+
+    class FakeLiberoEnv:
+        def __init__(self, *, cfg, **kwargs):
+            del kwargs
+            captured_cfgs.append(cfg)
+
+    names = (
+        "rlinf",
+        "rlinf.envs",
+        "rlinf.envs.sim",
+        "rlinf.envs.sim.libero",
+        "rlinf.envs.sim.libero.libero_env",
+        "rlinf.envs.sim.libero.utils",
+        "rlinf.envs.sim.libero.utils.benchmark",
+    )
+    modules: dict[str, ModuleType] = {}
+    for name in names:
+        module = ModuleType(name)
+        if name not in (
+            "rlinf.envs.sim.libero.libero_env",
+            "rlinf.envs.sim.libero.utils.benchmark",
+        ):
+            module.__path__ = []  # mark intermediate nodes as packages
+        modules[name] = module
+    modules["rlinf.envs.sim.libero.libero_env"].LiberoEnv = FakeLiberoEnv
+    modules["rlinf.envs.sim.libero.utils.benchmark"].get_benchmark = (
+        lambda _suite_name: lambda: suite
+    )
+    modules["rlinf.envs.sim.libero.utils"].benchmark = modules[
+        "rlinf.envs.sim.libero.utils.benchmark"
+    ]
+    for name, module in modules.items():
+        monkeypatch.setitem(sys.modules, name, module)
+    return FakeLiberoEnv, captured_cfgs
+
+
+def test_make_env_fails_loudly_on_empty_init_states(monkeypatch):
+    counts = [50] * 10
+    counts[2] = 0  # libero_10_task#2 ships 0 states (RLinf/RPent#246)
+    _install_fake_rlinf(monkeypatch, FakeSuite(counts))
+
+    with pytest.raises(RuntimeError, match="libero_10_task") as excinfo:
+        env_server.make_env(task_id=2, seed=0, suite_name="libero_10_task")
+
+    message = str(excinfo.value)
+    assert "task 2" in message
+    assert "0 init states" in message
+
+
+def test_make_env_error_lists_every_empty_task_in_suite(monkeypatch):
+    counts = [50] * 10
+    counts[3] = 0
+    counts[7] = 0  # libero_spatial_task#3/#7 ship 0 states (RLinf/RPent#246)
+    _install_fake_rlinf(monkeypatch, FakeSuite(counts))
+
+    with pytest.raises(RuntimeError, match=r"\[3, 7\]"):
+        env_server.make_env(task_id=3, seed=0, suite_name="libero_spatial_task")
+
+
+def test_make_env_keeps_offset_reset_id_for_healthy_tasks(monkeypatch):
+    counts = [10, 20, 30]
+    env_cls, captured_cfgs = _install_fake_rlinf(monkeypatch, FakeSuite(counts))
+
+    env = env_server.make_env(task_id=2, seed=5, suite_name="libero_spatial")
+
+    assert isinstance(env, env_cls)
+    # first_id = 10 + 20, rid = first_id + seed % trials = 30 + 5 % 30.
+    assert captured_cfgs[0].specific_reset_id == 35