Merge pull request #19 from opendilab/gl-dev

update v0.3.4

.github/workflows/doc.yml

@@ -4,7 +4,7 @@ on:
   push:
     branches: [main, 'doc/*']
   pull_request:
-    branches: [main, 'doc/*']
+    branches: [main]
 
 jobs:
   doc:
@@ -22,8 +22,7 @@ jobs:
       - name: Generate
         run: |
           python -m pip install .[doc]
-          wget http://opendilab.org/download/DI-drive/carla-0.9.9-py3.7-linux-x86_64.egg
-          easy_install carla-0.9.9-py3.7-linux-x86_64.egg
+          python -m pip install carla
           make -C ./docs html
           mv docs/build/html public
           rm -rf docs/build

.github/workflows/style.yml

@@ -7,15 +7,19 @@ jobs:
     runs-on: ubuntu-latest
     strategy:
       matrix:
-        python-version: [3.6, 3.7]
+        python-version: [3.7]
 
     steps:
       - uses: actions/checkout@v2
       - name: Set up Python ${{ matrix.python-version }}
         uses: actions/setup-python@v2
         with:
           python-version: ${{ matrix.python-version }}
-      - name: code style
+      - name: Install
+        run: |
+          python -m pip install .
+          python -c 'import core'
+      - name: Code style
         run: |
           python -m pip install .[style]
           bash format.sh ./core --test

CHANGELOG

@@ -1,3 +1,11 @@
+## v0.3.4 (2022.7.25)
+- Add Traj SAC and PPO policy
+- Add MetaDrive evaluator
+- Add MetaDrive TREX demo and doc
+- Update Carla support to 0.9.10
+- Fix bug of MetaDrive reset conflice with EnvManager
+- Fix bug in DingEnvManager
+
 ## v0.3.3 (2022.6.5)
 - Update readme, reorganize info
 - Add MetaDriveTrajEnv and doc

README.md

@@ -59,7 +59,7 @@ pip install -e .
 
 DI-engine and Pytorch will be installed automatically.
 
-In addition, at least one simulator in Carla and MetaDrive need to be able to run in DI-drive. [MetaDrive](https://decisionforce.github.io/metadrive/) can be easily installed via `pip`.
+In addition, at least one simulator in **Carla** and **MetaDrive** need to be installed to run in DI-drive. [MetaDrive](https://decisionforce.github.io/metadrive/) can be easily installed via `pip`.
 If [Carla](http://carla.org) server is used for simulation, users need to install 'Carla Python API' in addition. You can use either one of them or both. Make sure to modify the activated simulators in `core.__init__.py` to avoid import error.
 
 Please refer to the [installation guide](https://opendilab.github.io/DI-drive/installation/index.html) for details about the installation of **DI-drive**.
@@ -101,6 +101,10 @@ We provide detail guidance for IL and RL experiments in all simulators and quick
 - [Latent DRL](https://arxiv.org/abs/2001.08726)
 - MetaDrive Macro RL
 
+### Other Method
+
+- [DREX](http://proceedings.mlr.press/v100/brown20a.html)
+
 ## DI-drive Casezoo
 
 **DI-drive Casezoo** is a scenario set for training and testing the Autonomous Driving policy in simulator.
@@ -155,6 +159,7 @@ DI-drive
 |   |   |-- vae_model.py
 |   |   |-- vehicle_controller.py
 |   |-- policy
+|   |   |-- traj_policy
 |   |   |-- auto_policy.py
 |   |   |-- base_carla_policy.py
 |   |   |-- cilrs_policy.py

core/__init__.py

@@ -1,5 +1,5 @@
 __TITLE__ = "DI-drive"
-__VERSION__ = "0.3.3"
+__VERSION__ = "0.3.4"
 __DESCRIPTION__ = "Decision AI Auto-Driving Platform"
 __AUTHOR__ = "OpenDILab Contributors"
 __AUTHOR_EMAIL__ = "[email protected]"

core/envs/__init__.py

@@ -23,10 +23,12 @@
 if 'metadrive' in SIMULATORS:
     from .md_macro_env import MetaDriveMacroEnv
     from .md_traj_env import MetaDriveTrajEnv
-    env_map.update({
-        "Macro-v1": 'core.envs.md_macro_env:MetaDriveMacroEnv',
-        "Traj-v1": 'core.envs.md_traj_env:MetaDriveTrajEnv'
-    })
+    env_map.update(
+        {
+            "Macro-v1": 'core.envs.md_macro_env:MetaDriveMacroEnv',
+            "Traj-v1": 'core.envs.md_traj_env:MetaDriveTrajEnv'
+        }
+    )
 
 for k, v in env_map.items():
     if k not in registry.env_specs:

core/envs/drive_env_wrapper.py

@@ -52,6 +52,9 @@ def reset(self, *args, **kwargs) -> Any:
         obs = to_ndarray(obs, dtype=np.float32)
         if isinstance(obs, np.ndarray) and len(obs.shape) == 3:
             obs = obs.transpose((2, 0, 1))
+        elif isinstance(obs, dict):
+            birdview = obs['birdview'].transpose((2, 0, 1))
+            obs = {'birdview': birdview, 'vehicle_state': obs['vehicle_state']}
         self._final_eval_reward = 0.0
         return obs
 
@@ -75,6 +78,9 @@ def step(self, action: Any = None) -> BaseEnvTimestep:
         obs = to_ndarray(obs, dtype=np.float32)
         if isinstance(obs, np.ndarray) and len(obs.shape) == 3:
             obs = obs.transpose((2, 0, 1))
+        elif isinstance(obs, dict):
+            birdview = obs['birdview'].transpose((2, 0, 1))
+            obs = {'birdview': birdview, 'vehicle_state': obs['vehicle_state']}
         rew = to_ndarray([rew], dtype=np.float32)
         if done:
             info['final_eval_reward'] = self._final_eval_reward

core/envs/md_traj_env.py

@@ -11,7 +11,8 @@
 from ding.utils import ENV_REGISTRY
 from core.utils.simulator_utils.md_utils.discrete_policy import DiscreteMetaAction
 from core.utils.simulator_utils.md_utils.agent_manager_utils import TrajAgentManager
-from core.utils.simulator_utils.md_utils.engine_utils import TrajEngine
+from core.utils.simulator_utils.md_utils.engine_utils import TrajEngine, initialize_engine, close_engine, \
+     engine_initialized, set_global_random_seed
 from core.utils.simulator_utils.md_utils.traffic_manager_utils import TrafficMode
 from metadrive.constants import RENDER_MODE_NONE, DEFAULT_AGENT, REPLAY_DONE, TerminationState
 from metadrive.envs.base_env import BaseEnv
@@ -23,7 +24,6 @@
 from metadrive.envs.base_env import BASE_DEFAULT_CONFIG
 from metadrive.obs.top_down_obs_multi_channel import TopDownMultiChannel
 from metadrive.engine.base_engine import BaseEngine
-from metadrive.utils.utils import auto_termination
 
 DIDRIVE_DEFAULT_CONFIG = dict(
     # ===== Generalization =====
@@ -77,8 +77,8 @@
     # ===== Reward Scheme =====
     # See: https://github.com/decisionforce/metadrive/issues/283
     success_reward=10.0,  # 10.0,
-    out_of_road_penalty=1.0,  # 5.0,
-    crash_vehicle_penalty=1.0,  # 1.0,
+    out_of_road_penalty=5.0,  # 5.0,
+    crash_vehicle_penalty=5.0,  # 1.0,
     crash_object_penalty=5.0,  # 5.0,
     run_out_of_time_penalty=5.0,  # 5.0,
 
@@ -668,14 +668,14 @@ def _get_step_return(self, actions, engine_info):
 
         should_done = engine_info.get(REPLAY_DONE, False
                                       ) or (self.config["horizon"] and self.episode_steps >= self.config["horizon"])
-        termination_infos = self.for_each_vehicle(auto_termination, should_done)
+        #termination_infos = self.for_each_vehicle(self.auto_termination, should_done)
 
         step_infos = concat_step_infos([
             engine_info,
             done_infos,
             reward_infos,
             cost_infos,
-            termination_infos,
+            #termination_infos,
         ])
 
         if should_done:
@@ -847,10 +847,9 @@ def lazy_init(self):
         :return: None
         """
         # It is the true init() func to create the main vehicle and its module, to avoid incompatible with ray
-        if TrajEngine.singleton is not None:
+        if engine_initialized():
             return
-        TrajEngine.singleton = TrajEngine(self.config)
-        self.engine = TrajEngine.singleton
+        self.engine = initialize_engine(self.config)
         # engine setup
         self.setup_engine()
         # other optional initialization
@@ -883,3 +882,9 @@ def get_episode_max_step(self, distance, average_speed=6.5):
                                       ) * average_speed * self.config['physics_world_step_size']
         max_step = int(distance / average_dist_per_step) + 1
         return max_step
+
+    def close(self):
+        if self.engine is not None:
+            close_engine()
+        if self._top_down_renderer is not None:
+            self._top_down_renderer.close()