From 8be0ba67a66eb22c0148e524b14a97eab1dc7840 Mon Sep 17 00:00:00 2001
From: Aditya kumar singh <143548997+Adityakk9031@users.noreply.github.com>
Date: Thu, 13 Aug 2026 01:03:06 +0530
Subject: [PATCH 1/2] fix(controller): handle empty action chunks gracefully by
 raising PolicyError

---
 src/inspect_robots/controller.py |  7 +++++
 tests/test_controller.py         | 50 ++++++++++++++++++++++++++++++++
 2 files changed, 57 insertions(+)

diff --git a/src/inspect_robots/controller.py b/src/inspect_robots/controller.py
index 2acb742c..55cd00ed 100644
--- a/src/inspect_robots/controller.py
+++ b/src/inspect_robots/controller.py
@@ -23,6 +23,7 @@
 
 import numpy as np
 
+from inspect_robots.errors import PolicyError
 from inspect_robots.policy import Policy
 from inspect_robots.spaces import Box
 from inspect_robots.types import Action, Observation
@@ -58,6 +59,8 @@ def next_action(
         buffer: deque[Action] = store.setdefault(_BUFFER_KEY, deque())
         if not buffer:
             chunk = policy.act(observation)
+            if not chunk.actions:
+                raise PolicyError("policy emitted an empty ActionChunk (0 actions)")
             take = self.replan_interval or len(chunk)
             taken = list(chunk.actions)[:take]
             buffer.extend(taken)
@@ -163,6 +166,8 @@ def next_action(
     ) -> Action:
         """Query once for step ``t`` and blend every retained prediction for that step."""
         chunk = policy.act(observation)
+        if not chunk.actions:
+            raise PolicyError("policy emitted an empty ActionChunk (0 actions)")
         store.setdefault(_INFER_KEY, []).append((chunk.inference_latency_s, len(chunk)))
 
         buffer: list[tuple[int, list[Any], dict[str, Any]]] = store.setdefault(_ENSEMBLE_KEY, [])
@@ -179,6 +184,8 @@ def next_action(
         buffer.sort(key=lambda e: e[0])
 
         predictions = [acts[t - q] for (q, acts, _meta) in buffer]
+        if not predictions:
+            raise PolicyError("no valid action predictions available for current step")
         weights = np.exp(-self.m * np.arange(len(predictions)))
         weights = weights / weights.sum()
         blended = np.average(np.stack(predictions), axis=0, weights=weights)
diff --git a/tests/test_controller.py b/tests/test_controller.py
index 29a8bbc7..24a481e0 100644
--- a/tests/test_controller.py
+++ b/tests/test_controller.py
@@ -99,3 +99,53 @@ def test_stacked_smoothing_controllers_keep_separate_state() -> None:
         expected.append(outer_prev)
 
     assert np.allclose(actual, expected)
+
+
+class _EmptyChunkPolicy:
+    """Emits an empty ActionChunk with zero actions."""
+
+    def __init__(self) -> None:
+        self.info = PolicyInfo(
+            name="empty",
+            action_space=Box(shape=(2,), semantics=ActionSemantics(control_mode="joint_delta")),
+        )
+        self.config = PolicyConfig()
+
+    def reset(self, scene: Scene) -> None:
+        pass
+
+    def act(self, observation: Observation) -> ActionChunk:
+        chunk = ActionChunk(actions=[Action(data=np.zeros(2))])
+        object.__setattr__(chunk, "actions", [])
+        return chunk
+
+
+def test_default_controller_raises_policy_error_on_empty_chunk() -> None:
+    import pytest
+
+    from inspect_robots.errors import PolicyError
+
+    policy = _EmptyChunkPolicy()
+    embodiment = CubePickEmbodiment()
+    obs = _obs(embodiment)
+    store: dict[str, object] = {}
+    ctrl = DefaultController()
+
+    with pytest.raises(PolicyError, match="empty ActionChunk"):
+        ctrl.next_action(policy, obs, 0, store)
+
+
+def test_ensembling_controller_raises_policy_error_on_empty_chunk() -> None:
+    import pytest
+
+    from inspect_robots.controller import EnsemblingController
+    from inspect_robots.errors import PolicyError
+
+    policy = _EmptyChunkPolicy()
+    embodiment = CubePickEmbodiment()
+    obs = _obs(embodiment)
+    store: dict[str, object] = {}
+    ctrl = EnsemblingController(policy.info.action_space)
+
+    with pytest.raises(PolicyError, match="empty ActionChunk"):
+        ctrl.next_action(policy, obs, 0, store)

From 52d9f5315356bda16a92ade5b5bf8fb87c9fc399 Mon Sep 17 00:00:00 2001
From: Aditya kumar singh <143548997+Adityakk9031@users.noreply.github.com>
Date: Thu, 13 Aug 2026 01:16:24 +0530
Subject: [PATCH 2/2] fix(controller): add pragma no cover for defensive
 unreachable predictions check

---
 src/inspect_robots/controller.py | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/src/inspect_robots/controller.py b/src/inspect_robots/controller.py
index 55cd00ed..814e56f9 100644
--- a/src/inspect_robots/controller.py
+++ b/src/inspect_robots/controller.py
@@ -184,8 +184,9 @@ def next_action(
         buffer.sort(key=lambda e: e[0])
 
         predictions = [acts[t - q] for (q, acts, _meta) in buffer]
-        if not predictions:
+        if not predictions:  # pragma: no cover
             raise PolicyError("no valid action predictions available for current step")
+
         weights = np.exp(-self.m * np.arange(len(predictions)))
         weights = weights / weights.sum()
         blended = np.average(np.stack(predictions), axis=0, weights=weights)