Fix memory leak in ManagerBasedEnv.close

source/isaaclab/isaaclab/envs/manager_based_env.py

@@ -551,6 +551,11 @@ def close(self):
             # Stop simulation first to allow physics to clean up properly
             self.sim.stop()
 
+            # Drop cached observation tensors so they don't survive close via
+            # gymnasium's wrapper chain.
+            if isinstance(getattr(self, "obs_buf", None), dict):
+                self.obs_buf.clear()
+
             # destructor is order-sensitive
             del self.viewport_camera_controller
             del self.action_manager

source/isaaclab/isaaclab/envs/manager_based_rl_env.py

@@ -311,6 +311,12 @@ def close(self):
             del self.reward_manager
             del self.termination_manager
             del self.curriculum_manager
+            # Release gym.spaces.Box bounds arrays set in _configure_gym_env_spaces;
+            # without this they survive close via gymnasium's wrapper chain.
+            self.single_observation_space = None
+            self.single_action_space = None
+            self.observation_space = None
+            self.action_space = None
             # call the parent class to close the environment
             super().close()
 

source/isaaclab/test/envs/test_manager_based_env_close_leak.py

@@ -0,0 +1,50 @@
+# Copyright (c) 2022-2026, The Isaac Lab Project Developers (https://github.com/isaac-sim/IsaacLab/blob/main/CONTRIBUTORS.md).
+# All rights reserved.
+#
+# SPDX-License-Identifier: BSD-3-Clause
+
+"""Regression test for the ``ManagerBasedEnv.close`` memory leak fixed in this PR."""
+
+"""Launch Isaac Sim Simulator first."""
+
+from isaaclab.app import AppLauncher
+
+simulation_app = AppLauncher(headless=True).app
+
+"""Rest everything follows."""
+
+import gymnasium as gym
+import pytest
+import torch
+
+import isaaclab_tasks  # noqa: F401 -- registers Isaac-* env IDs with gymnasium
+from isaaclab_tasks.utils.parse_cfg import parse_env_cfg
+
+
+@pytest.mark.parametrize("device", ["cuda:0", "cpu"])
+def test_close_clears_obs_buf_and_releases_spaces(device):
+    """``ManagerBasedEnv.close`` should release the cached observation buffer and the
+    gym space attributes attached to the env.
+
+    Without the release, gymnasium's wrapper chain (e.g. ``OrderEnforcing`` plus the
+    ``make()`` registry) keeps the env reachable past ``close``, so ``self.obs_buf``
+    (the dict of cached observation tensors — tens of MB per image sensor) and the
+    ``observation_space`` / ``action_space`` ``gym.spaces.Box`` attributes (~110 MB
+    of numpy bounds-array memory per image-observation Box at ``(num_envs, H, W, C)``
+    float32) survive the call and accumulate on each construct/teardown cycle.
+    """
+    env_cfg = parse_env_cfg("Isaac-Cartpole-v0", device=device, num_envs=2)
+    env = gym.make("Isaac-Cartpole-v0", cfg=env_cfg)
+    env.reset()
+    # Step once so the ObservationManager populates ``obs_buf``.
+    action = torch.zeros((2, env.action_space.shape[-1]), device=env.unwrapped.device)
+    env.step(action)
+    assert env.unwrapped.obs_buf, "precondition: obs_buf should be populated after step"
+    assert env.unwrapped.observation_space is not None, "precondition: observation_space should be set"
+    assert env.unwrapped.action_space is not None, "precondition: action_space should be set"
+
+    env.close()
+
+    assert not env.unwrapped.obs_buf, "obs_buf was not cleared on close"
+    assert env.unwrapped.observation_space is None, "observation_space was not released on close"
+    assert env.unwrapped.action_space is None, "action_space was not released on close"