Keep the scaled action space valid for a negative scale

supersuit/generic_wrappers/basic_wrappers.py

@@ -49,9 +49,11 @@ def normalize_obs_v0(env, env_min=0.0, env_max=1.0):
 def clip_actions_v0(env):
     return action_lambda_v1(
         env,
-        lambda action, act_space: np.clip(action, act_space.low, act_space.high)
-        if action is not None
-        else None,
+        lambda action, act_space: (
+            np.clip(action, act_space.low, act_space.high)
+            if action is not None
+            else None
+        ),
         lambda act_space: act_space,
     )
 
@@ -61,13 +63,22 @@ def change_act_space(act_space):
         assert isinstance(
             act_space, Box
         ), "scale_actions_v0 only works with a Box action space"
-        return Box(low=act_space.low * scale, high=act_space.high * scale)
+        # A negative scale maps [low, high] onto [high * scale, low * scale], so the
+        # two bounds swap places. Multiplying them in place left a Box whose low
+        # exceeded its high, and Gymnasium rejected it from inside its own
+        # constructor without naming this wrapper or its `scale` argument.
+        scaled_low = act_space.low * scale
+        scaled_high = act_space.high * scale
+        return Box(
+            low=np.minimum(scaled_low, scaled_high),
+            high=np.maximum(scaled_low, scaled_high),
+        )
 
     return action_lambda_v1(
         env,
-        lambda action, act_space: np.asarray(action) * scale
-        if action is not None
-        else None,
+        lambda action, act_space: (
+            np.asarray(action) * scale if action is not None else None
+        ),
         lambda act_space: change_act_space(act_space),
     )
 

test/aec_mock_test.py

@@ -444,3 +444,33 @@ def test_scale_action_wrapper():
         base_act_spaces = {f"a{idx}": Discrete(5) for idx in range(2)}
         base_env = DummyScaleEnv(base_obs, base_obs_space, base_act_spaces)
         wrapped_env = scale_actions_v0(base_env, 2)
+
+
+def test_scale_action_wrapper_negative_scale():
+    """A negative scale mirrors the interval, so the bounds have to swap.
+
+    `Box(low=low * scale, high=high * scale)` left low above high for a negative
+    scale, and Gymnasium rejected the space from inside its own constructor
+    without naming this wrapper or its `scale` argument.
+    """
+    base_obs = {f"a{idx}": np.zeros([3], dtype=np.float32) for idx in range(2)}
+    base_obs_space = {
+        f"a{idx}": Box(low=np.float32(0.0), high=np.float32(10.0), shape=[3])
+        for idx in range(2)
+    }
+    base_act_spaces = {
+        f"a{idx}": Box(low=np.float32(1.0), high=np.float32(5.0), shape=[3])
+        for idx in range(2)
+    }
+
+    base_env = DummyScaleEnv(base_obs, base_obs_space, base_act_spaces)
+    wrapped_env = scale_actions_v0(base_env, -2.0)
+
+    space = wrapped_env.action_space(wrapped_env.possible_agents[0])
+    assert np.allclose(space.low, -10.0)
+    assert np.allclose(space.high, -2.0)
+
+    wrapped_env.reset()
+    wrapped_env.step(np.array([2, 1, 3], dtype=np.float32))
+    scaled_action = wrapped_env.observe(wrapped_env.agents[0])
+    assert (scaled_action == np.array([-4, -2, -6], dtype=np.float32)).all()