[Feature] IsaacLab wrapper

ghstack-source-id: 66fe16a
Pull-Request-resolved: #2937

Author: vmoens

.github/unittest/linux_libs/scripts_gym/setup_env.sh

@@ -10,7 +10,6 @@ set -e
 this_dir="$( cd "$( dirname "${BASH_SOURCE[0]}" )" >/dev/null 2>&1 && pwd )"
 # Avoid error: "fatal: unsafe repository"
 apt-get update && apt-get install -y git wget gcc g++
-
 apt-get install -y libglfw3 libgl1-mesa-glx libosmesa6 libglew-dev libsdl2-dev libsdl2-2.0-0
 apt-get install -y libglvnd0 libgl1 libglx0 libegl1 libgles2 xvfb libegl-dev libx11-dev freeglut3-dev
 

.github/unittest/linux_libs/scripts_isaaclab/isaac.sh

@@ -0,0 +1,80 @@
+#!/usr/bin/env bash
+
+set -e
+set -v
+
+#if [[ "${{ github.ref }}" =~ release/* ]]; then
+#  export RELEASE=1
+#  export TORCH_VERSION=stable
+#else
+export RELEASE=0
+export TORCH_VERSION=nightly
+#fi
+
+set -euo pipefail
+export PYTHON_VERSION="3.10"
+export CU_VERSION="12.8"
+export TAR_OPTIONS="--no-same-owner"
+export UPLOAD_CHANNEL="nightly"
+export TF_CPP_MIN_LOG_LEVEL=0
+export BATCHED_PIPE_TIMEOUT=60
+export TD_GET_DEFAULTS_TO_NONE=1
+
+nvidia-smi
+
+# Setup
+apt-get update && apt-get install -y git wget gcc g++
+apt-get install -y libglfw3 libgl1-mesa-glx libosmesa6 libglew-dev libsdl2-dev libsdl2-2.0-0
+apt-get install -y libglvnd0 libgl1 libglx0 libegl1 libgles2 xvfb libegl-dev libx11-dev freeglut3-dev
+
+git config --global --add safe.directory '*'
+root_dir="$(git rev-parse --show-toplevel)"
+conda_dir="${root_dir}/conda"
+env_dir="${root_dir}/env"
+lib_dir="${env_dir}/lib"
+
+cd "${root_dir}"
+
+# install conda
+printf "* Installing conda\n"
+wget -O miniconda.sh "http://repo.continuum.io/miniconda/Miniconda3-latest-Linux-x86_64.sh"
+bash ./miniconda.sh -b -f -p "${conda_dir}"
+eval "$(${conda_dir}/bin/conda shell.bash hook)"
+
+
+conda create -n env_isaaclab python=3.10 -y
+conda activate env_isaaclab
+
+# Pin pytorch to 2.5.1 for IsaacLab
+conda install pytorch==2.5.1 torchvision==0.20.1 pytorch-cuda=12.4 -c pytorch -c nvidia -y
+
+pip install --upgrade pip
+pip install 'isaacsim[all,extscache]==4.5.0' --extra-index-url https://pypi.nvidia.com
+conda install conda-forge::"cmake>3.22" -y
+
+git clone https://github.com/isaac-sim/IsaacLab.git
+cd IsaacLab
+conda run -p ${conda_dir} ./isaaclab.sh --install sb3
+cd ../
+
+# install tensordict
+if [[ "$RELEASE" == 0 ]]; then
+  conda install "anaconda::cmake>=3.22" -y
+  conda run -p ${conda_dir} python3 -m pip install "pybind11[global]"
+  conda run -p ${conda_dir} python3 -m pip install git+https://github.com/pytorch/tensordict.git
+else
+  conda run -p ${conda_dir} python3 -m pip install tensordict
+fi
+
+# smoke test
+conda run -p ${conda_dir} python -c "import tensordict"
+
+printf "* Installing torchrl\n"
+conda run -p ${conda_dir} python setup.py develop
+conda run -p ${conda_dir} python -c "import torchrl"
+
+# Install pytest
+conda run -p ${conda_dir} python -m pip install pytest pytest-cov pytest-mock pytest-instafail pytest-rerunfailures pytest-error-for-skips pytest-asyncio
+
+# Run tests
+conda run -p ${conda_dir} python -m pytest test/test_libs.py -k isaac

.github/workflows/test-linux-libs.yml

@@ -230,6 +230,24 @@ jobs:
         ./.github/unittest/linux_libs/scripts_gym/batch_scripts.sh
         ./.github/unittest/linux_libs/scripts_gym/post_process.sh
 
+  unittests-isaaclab:
+    strategy:
+      matrix:
+        python_version: ["3.10"]
+        cuda_arch_version: ["12.8"]
+    if: ${{ github.event_name == 'push' || contains(github.event.pull_request.labels.*.name, 'Environments/Isaac') }}
+    uses: vmoens/test-infra/.github/workflows/isaac_linux_job_v2.yml@main
+    with:
+      repository: pytorch/rl
+      runner: "linux.g5.4xlarge.nvidia.gpu"
+      docker-image: "nvcr.io/nvidia/isaac-lab:2.1.0"
+      gpu-arch-type: cuda
+      gpu-arch-version: "12.8"
+      timeout: 120
+      test-infra-repository: vmoens/test-infra
+      script: |
+        ./.github/unittest/linux_libs/scripts_isaaclab/isaac.sh
+
   unittests-jumanji:
     strategy:
       matrix:

docs/source/reference/envs.rst

@@ -1417,6 +1417,7 @@ the following function will return ``1`` when queried:
     HabitatEnv
     IsaacGymEnv
     IsaacGymWrapper
+    IsaacLabWrapper
     JumanjiEnv
     JumanjiWrapper
     MeltingpotEnv

test/test_libs.py

@@ -32,32 +32,6 @@
 import pytest
 import torch
 
-if os.getenv("PYTORCH_TEST_FBCODE"):
-    from pytorch.rl.test._utils_internal import (
-        _make_multithreaded_env,
-        CARTPOLE_VERSIONED,
-        get_available_devices,
-        get_default_devices,
-        HALFCHEETAH_VERSIONED,
-        PENDULUM_VERSIONED,
-        PONG_VERSIONED,
-        rand_reset,
-        retry,
-        rollout_consistency_assertion,
-    )
-else:
-    from _utils_internal import (
-        _make_multithreaded_env,
-        CARTPOLE_VERSIONED,
-        get_available_devices,
-        get_default_devices,
-        HALFCHEETAH_VERSIONED,
-        PENDULUM_VERSIONED,
-        PONG_VERSIONED,
-        rand_reset,
-        retry,
-        rollout_consistency_assertion,
-    )
 from packaging import version
 from tensordict import (
     assert_allclose_td,
@@ -155,6 +129,33 @@
     ValueOperator,
 )
 
+if os.getenv("PYTORCH_TEST_FBCODE"):
+    from pytorch.rl.test._utils_internal import (
+        _make_multithreaded_env,
+        CARTPOLE_VERSIONED,
+        get_available_devices,
+        get_default_devices,
+        HALFCHEETAH_VERSIONED,
+        PENDULUM_VERSIONED,
+        PONG_VERSIONED,
+        rand_reset,
+        retry,
+        rollout_consistency_assertion,
+    )
+else:
+    from _utils_internal import (
+        _make_multithreaded_env,
+        CARTPOLE_VERSIONED,
+        get_available_devices,
+        get_default_devices,
+        HALFCHEETAH_VERSIONED,
+        PENDULUM_VERSIONED,
+        PONG_VERSIONED,
+        rand_reset,
+        retry,
+        rollout_consistency_assertion,
+    )
+
 _has_d4rl = importlib.util.find_spec("d4rl") is not None
 
 _has_mo = importlib.util.find_spec("mo_gymnasium") is not None
@@ -166,6 +167,9 @@
 _has_minari = importlib.util.find_spec("minari") is not None
 
 _has_gymnasium = importlib.util.find_spec("gymnasium") is not None
+
+_has_isaaclab = importlib.util.find_spec("scripts_isaaclab") is not None
+
 _has_gym_regular = importlib.util.find_spec("gym") is not None
 if _has_gymnasium:
     set_gym_backend("gymnasium").set()
@@ -4541,6 +4545,65 @@ def test_render(self, rollout_steps):
         assert not torch.equal(rollout_penultimate_image, image_from_env)
 
 
+@pytest.mark.skipif(not _has_isaaclab, reason="Isaaclab not found")
+class TestIsaacLab:
+    @pytest.fixture(scope="class")
+    def env(self):
+        torch.manual_seed(0)
+        import argparse
+
+        # This code block ensures that the Isaac app is started in headless mode
+        from isaaclab.app import AppLauncher
+
+        parser = argparse.ArgumentParser(description="Train an RL agent with TorchRL.")
+        AppLauncher.add_app_launcher_args(parser)
+        args_cli, hydra_args = parser.parse_known_args(["--headless"])
+        AppLauncher(args_cli)
+
+        # Imports and env
+        import gymnasium as gym
+        import isaaclab_tasks  # noqa: F401
+        from isaaclab_tasks.manager_based.classic.ant.ant_env_cfg import AntEnvCfg
+        from torchrl.envs.libs.isaac_lab import IsaacLabWrapper
+
+        torchrl_logger.info("Making IsaacLab env...")
+        env = gym.make("Isaac-Ant-v0", cfg=AntEnvCfg())
+        torchrl_logger.info("Wrapping IsaacLab env...")
+        try:
+            env = IsaacLabWrapper(env)
+            yield env
+        finally:
+            torchrl_logger.info("Closing IsaacLab env...")
+            env.close()
+            torchrl_logger.info("Closed")
+
+    def test_isaaclab(self, env):
+        assert env.batch_size == (4096,)
+        assert env._is_batched
+        torchrl_logger.info("Checking env specs...")
+        env.check_env_specs(break_when_any_done="both")
+        torchrl_logger.info("Check succeeded!")
+
+    def test_isaac_collector(self, env):
+        col = SyncDataCollector(
+            env, env.rand_action, frames_per_batch=1000, total_frames=100_000_000
+        )
+        try:
+            for data in col:
+                assert data.shape == (4096, 1)
+                break
+        finally:
+            # We must do that, otherwise `__del__` calls `shutdown` and the next test will fail
+            col.shutdown(close_env=False)
+
+    def test_isaaclab_reset(self, env):
+        # Make a rollout that will stop as soon as a trajectory reaches a done state
+        r = env.rollout(1_000_000)
+
+        # Check that done obs are None
+        assert not r["next", "policy"][r["next", "done"].squeeze(-1)].isfinite().any()
+
+
 if __name__ == "__main__":
     args, unknown = argparse.ArgumentParser().parse_known_args()
     pytest.main([__file__, "--capture", "no", "--exitfirst"] + unknown)

torchrl/collectors/collectors.py

@@ -272,15 +272,19 @@ def pause(self):
             f"Collector pause() is not implemented for {type(self).__name__}."
         )
 
-    def async_shutdown(self, timeout: float | None = None) -> None:
+    def async_shutdown(
+        self, timeout: float | None = None, close_env: bool = True
+    ) -> None:
         """Shuts down the collector when started asynchronously with the `start` method.
 
         Arg:
             timeout (float, optional): The maximum time to wait for the collector to shutdown.
+            close_env (bool, optional): If True, the collector will close the contained environment.
+                Defaults to `True`.
 
         .. seealso:: :meth:`~.start`
         """
-        return self.shutdown(timeout=timeout)
+        return self.shutdown(timeout=timeout, close_env=close_env)
 
     def update_policy_weights_(
         self,
@@ -336,7 +340,7 @@ def next(self):
             return None
 
     @abc.abstractmethod
-    def shutdown(self, timeout: float | None = None) -> None:
+    def shutdown(self, timeout: float | None = None, close_env: bool = True) -> None:
         raise NotImplementedError
 
     @abc.abstractmethod
@@ -1311,12 +1315,14 @@ def _run_iterator(self):
             if self._stop:
                 return
 
-    def async_shutdown(self, timeout: float | None = None) -> None:
+    def async_shutdown(
+        self, timeout: float | None = None, close_env: bool = True
+    ) -> None:
         """Finishes processes started by ray.init() during async execution."""
         self._stop = True
         if hasattr(self, "_thread") and self._thread.is_alive():
             self._thread.join(timeout=timeout)
-        self.shutdown()
+        self.shutdown(close_env=close_env)
 
     def _postproc(self, tensordict_out):
         if self.split_trajs:
@@ -1576,14 +1582,20 @@ def reset(self, index=None, **kwargs) -> None:
         )
         self._shuttle["collector"] = collector_metadata
 
-    def shutdown(self, timeout: float | None = None) -> None:
-        """Shuts down all workers and/or closes the local environment."""
+    def shutdown(self, timeout: float | None = None, close_env: bool = True) -> None:
+        """Shuts down all workers and/or closes the local environment.
+
+        Args:
+            timeout (float, optional): The timeout for closing pipes between workers.
+                No effect for this class.
+            close_env (bool, optional): Whether to close the environment. Defaults to `True`.
+        """
         if not self.closed:
             self.closed = True
             del self._shuttle
             if self._use_buffers:
                 del self._final_rollout
-            if not self.env.is_closed:
+            if close_env and not self.env.is_closed:
                 self.env.close()
             del self.env
         return
@@ -2385,8 +2397,17 @@ def __del__(self):
             # __del__ will not affect the program.
             pass
 
-    def shutdown(self, timeout: float | None = None) -> None:
-        """Shuts down all processes. This operation is irreversible."""
+    def shutdown(self, timeout: float | None = None, close_env: bool = True) -> None:
+        """Shuts down all processes. This operation is irreversible.
+
+        Args:
+            timeout (float, optional): The timeout for closing pipes between workers.
+            close_env (bool, optional): Whether to close the environment. Defaults to `True`.
+        """
+        if not close_env:
+            raise RuntimeError(
+                f"Cannot shutdown {type(self).__name__} collector without environment being closed."
+            )
         self._shutdown_main(timeout)
 
     def _shutdown_main(self, timeout: float | None = None) -> None:
@@ -2659,7 +2680,11 @@ def next(self):
         return super().next()
 
     # for RPC
-    def shutdown(self, timeout: float | None = None) -> None:
+    def shutdown(self, timeout: float | None = None, close_env: bool = True) -> None:
+        if not close_env:
+            raise RuntimeError(
+                f"Cannot shutdown {type(self).__name__} collector without environment being closed."
+            )
         if hasattr(self, "out_buffer"):
             del self.out_buffer
         if hasattr(self, "buffers"):
@@ -3032,9 +3057,13 @@ def next(self):
         return super().next()
 
     # for RPC
-    def shutdown(self, timeout: float | None = None) -> None:
+    def shutdown(self, timeout: float | None = None, close_env: bool = True) -> None:
         if hasattr(self, "out_tensordicts"):
             del self.out_tensordicts
+        if not close_env:
+            raise RuntimeError(
+                f"Cannot shutdown {type(self).__name__} collector without environment being closed."
+            )
         return super().shutdown(timeout=timeout)
 
     # for RPC
@@ -3376,8 +3405,8 @@ def next(self):
         return super().next()
 
     # for RPC
-    def shutdown(self, timeout: float | None = None) -> None:
-        return super().shutdown(timeout=timeout)
+    def shutdown(self, timeout: float | None = None, close_env: bool = True) -> None:
+        return super().shutdown(timeout=timeout, close_env=close_env)
 
     # for RPC
     def set_seed(self, seed: int, static_seed: bool = False) -> int: