Add NeurIPS2022 competition agent to agent zoo

CHANGELOG.md

@@ -10,6 +10,7 @@ Copy and pasting the git commit messages is __NOT__ enough.
 
 ## [Unreleased]
 ### Added
+- Added a zoo agent, named Interaction-aware Motion Prediction, from NeurIPS2022 submission. This zoo agent runs in benchmark `driving_smarts==0.0`.
 ### Changed
 ### Deprecated
 ### Fixed

baselines/driving_smarts/v0/interaction_aware_motion_prediction.yaml

@@ -0,0 +1,7 @@
+agent:
+  interface:
+    action_space: "TargetPose" # we can only allow `RelativeTargetPose` and `TargetPose`
+    img_meters: 50 # Observation image area size in meters.
+    img_pixels: 112 # Observation image size in pixels.
+    observation_options: "unformatted"
+  locator: "zoo.policies:interaction-aware-motion-prediction-agent-v0"

docs/benchmarks/benchmark.rst

@@ -1,7 +1,7 @@
 .. _benchmark:
 
-Driving SMARTS Competition Benchmark
-====================================
+Driving SMARTS Competition
+==========================
 
 The Driving SMARTS competition benchmark is a benchmark derived from the
 NeurIPS 2022 Driving SMARTS Competition.
@@ -30,7 +30,7 @@ available.
 
 .. code:: bash
 
-   > scl benchmark list 
+   $ scl benchmark list 
    BENCHMARK_NAME               BENCHMARK_ID             VERSIONS
    - Driving SMARTS:            driving_smarts           0.0 0.1
 
@@ -133,7 +133,7 @@ file.
 
 .. code:: bash
 
-   > scl benchmark run driving_smarts "./baselines/driving_smarts/v0/agent_config.yaml" --auto-install # --auto-install only needs to be used to get dependencies.
+   $ scl benchmark run driving_smarts "./baselines/driving_smarts/v0/agent_config.yaml" --auto-install # --auto-install only needs to be used to get dependencies.
    Starting `Driving SMARTS V1` benchmark.
    This is a cleaned up version of the Driving SMARTS benchmark.
 
@@ -151,17 +151,17 @@ file.
 
    `Driving SMARTS V0` result:
    - completion: 1
-   - humanness: 0.2
-   - rules: 0.8
-   - time: 0.3
-   - overall: 1.7
+   - humanness: 0.7
+   - rules: 0.9
+   - time: 0.2
+   - overall: 0.504
 
 A particular version of a benchmark can be targeted using a modified
 syntax ``benchmark_name==version``:
 
 .. code:: bash
 
-   > scl benchmark run driving_smarts==0.0 ...
+   $ scl benchmark run driving_smarts==0.0 "./baselines/driving_smarts/v0/agent_config.yaml" --auto-install
 
 Advanced Configuration
 ----------------------
@@ -178,7 +178,7 @@ what benchmarks are currently available. This can be passed using
 
 .. code:: bash
 
-   > scl benchmark run --benchmark-listing benchmark_listing.yaml driving_smarts "./baselines/driving_smarts/v0/agent_config.yaml"
+   $ scl benchmark run --benchmark-listing benchmark_listing.yaml driving_smarts "./baselines/driving_smarts/v0/agent_config.yaml"
 
 WARNING! Since with ``scl benchmark run`` this listing directs to a code
 ``entrypoint`` do not use this with a listing file from an unknown
@@ -192,7 +192,7 @@ file.
 
 .. code:: bash
 
-   > scl benchmark list --benchmark-listing benchmark_listing.yaml
+   $ scl benchmark list --benchmark-listing benchmark_listing.yaml
 
 Listing File
 ^^^^^^^^^^^^

docs/conf.py

@@ -46,11 +46,11 @@
 
 extlinks = {
     "examples": (
-        "https://github.com/huawei-noah/SMARTS/blob/develop/examples/%s",
+        "https://github.com/huawei-noah/SMARTS/blob/master/examples/%s",
         "%s",
     ),
     "scenarios": (
-        "https://github.com/huawei-noah/SMARTS/blob/develop/scenarios/%s",
+        "https://github.com/huawei-noah/SMARTS/blob/master/scenarios/%s",
         "%s",
     ),
 }
@@ -65,6 +65,7 @@
     "setup.py",
     "scenarios",
     "smarts/ros",
+    "zoo/policies/interaction_aware_motion_prediction",
 ]
 apidoc_extra_args = [
     "--force",
@@ -77,6 +78,7 @@
 autodoc_mock_imports = [
     "cpuinfo",
     "cv2",
+    "gymnasium",
     "lxml",
     "mdutils",
     "moviepy",
@@ -86,6 +88,7 @@
     "ray",
     "tabulate",
     "tools",
+    "torch",
     "waymo_open_dataset",
 ]
 

docs/ecosystem/agent_zoo.rst

@@ -0,0 +1,54 @@
+.. _agent_zoo:
+
+Agent Zoo
+=========
+
+Contributed agents are hosted at `Agent Zoo <https://github.com/huawei-noah/SMARTS/tree/master/zoo/policies>`_ .
+
+These agents should be installed prior to being used in SMARTS environments.
+
+.. code-block:: bash
+
+    $ cd <path>/SMARTS
+    # Install a zoo agent.
+    $ scl zoo install <agent>
+    # e.g. scl zoo install zoo/policies/interaction_aware_motion_prediction
+
+.. note::
+
+    To build a wheel, execute:
+
+    .. code-block:: bash
+
+        $ scl zoo build <agent>
+        # e.g. scl zoo build zoo/policies/interaction_aware_motion_prediction
+
+Agents
+------
+
+#. Interaction-aware motion prediction agent
+
+   * This agent was contributed as part of NeurIPS2022 Driving SMARTS competition.
+   * Agent `code <https://github.com/huawei-noah/SMARTS/tree/master/zoo/policies/interaction_aware_motion_prediction>`_.
+   * Run as follows:
+
+     .. code-block:: bash
+
+       $ cd <path>/SMARTS
+       # Install a zoo agent.
+       $ scl zoo install zoo/policies/interaction_aware_motion_prediction
+       $ scl benchmark run driving_smarts==0.0 "./baselines/driving_smarts/v0/interaction_aware_motion_prediction.yaml" --auto-install 
+
+Contribute agents
+-----------------
+
+.. note::
+    This section is only for contributing to the agent zoo.
+
+First, add the new agent to `Agent Zoo <https://github.com/huawei-noah/SMARTS/tree/master/zoo/policies>`_. It should contain 
+
+* package setup file,
+* inference code with trained model, and
+* algorithm explanation. 
+
+Then, register the newly added zoo agent in ``SMARTS/zoo/policies/__init__.py``.

docs/index.rst

@@ -78,6 +78,7 @@ If you use SMARTS in your research, please cite the `paper <https://arxiv.org/ab
    ecosystem/sumo.rst
    ecosystem/waymo.rst
    ecosystem/rllib.rst
+   ecosystem/agent_zoo.rst
 
 .. toctree::
    :hidden:

docs/resources/contributing.rst

@@ -54,7 +54,7 @@ Pre-Push Checklist
 ------------------
 
 1. Do your best to see that your code compiles locally.
-2. Do not push to ``master``. Instead make a branch and a pull request to the ``develop`` branch.
+2. Do not push directly to ``master``. Instead make a branch and a pull request to the ``master`` branch.
 
 Submission of a Pull Request
 ----------------------------

docs/setup.rst

@@ -6,7 +6,7 @@ Setup
 Prerequisites
 -------------
 
-+ python3 (3.7 and 3.8)
++ python3 (3.7 or 3.8)
 + ubuntu (>=16.04)
 
 Installation
@@ -20,15 +20,16 @@ Run the following commands to setup the SMARTS simulator.
     $ cd <path/to/SMARTS>
 
     # For latest stable release
-    $ git checkout master
+    $ git checkout tags/<tag_name>
+    # e.g., git checkout tags/v1.0.3
 
 .. note::
 
     Alternatively, to use the current development (i.e., bleeding edge) version:
 
     .. code-block:: bash
 
-        $ git checkout develop
+        $ git checkout master
 
 .. code-block:: bash
 

smarts/benchmark/entrypoints/benchmark_runner_v0.py

@@ -21,14 +21,13 @@
 # THE SOFTWARE.
 import logging
 import os
-from typing import Callable, Dict, Generator, List, Tuple
+from typing import List, Tuple
 
 import gymnasium as gym
 import psutil
 import ray
 
 from smarts.benchmark.driving_smarts import load_config
-from smarts.benchmark.driving_smarts.v0 import DEFAULT_CONFIG
 from smarts.core.utils.logging import suppress_output
 from smarts.env.gymnasium.wrappers.metrics import Metrics, Score
 from smarts.zoo import registry as agent_registry
@@ -53,18 +52,21 @@ def _eval_worker_local(name, env_config, episodes, agent_config, error_tolerant=
         **agent_config["interface"],
     )
     env = Metrics(env)
-    agent = agent_registry.make_agent(
-        locator=agent_config["locator"],
-        **agent_config["kwargs"],
-    )
+    agents = {
+        agent_id: agent_registry.make_agent(
+            locator=agent_config["locator"], **agent_config.get("kwargs", {})
+        )
+        for agent_id in env.agent_ids
+    }
 
     observation, info = env.reset()
     current_resets = 0
     try:
         while current_resets < episodes:
             try:
                 action = {
-                    agent_id: agent.act(obs) for agent_id, obs in observation.items()
+                    agent_id: agents[agent_id].act(obs)
+                    for agent_id, obs in observation.items()
                 }
                 # assert env.action_space.contains(action)
             except Exception:
@@ -185,7 +187,7 @@ def format_scores_total(named_scores: List[Tuple[str, Score]], scenario_count):
     print(format_one_line_scores(named_scores))
     print()
     print("`Driving SMARTS V0` averaged result:")
-    print(format_scores_total(named_scores, len(env_args) or 1))
+    print(format_scores_total(named_scores, len(env_args)))
 
 
 def benchmark_from_configs(benchmark_config, agent_config, debug_log=False):

smarts/env/gymnasium/driving_smarts_competition_env.py

@@ -24,7 +24,7 @@
 import os
 import pathlib
 from functools import partial
-from typing import Any, Dict, Optional, Tuple
+from typing import Any, Dict, Optional, Tuple, Union
 
 import gymnasium as gym
 import numpy as np
@@ -42,6 +42,7 @@
 )
 from smarts.core.controllers import ActionSpaceType
 from smarts.env.gymnasium.hiway_env_v1 import HiWayEnvV1, SumoOptions
+from smarts.env.utils.observation_conversion import ObservationOptions
 from smarts.sstudio.scenario_construction import build_scenario
 
 logger = logging.getLogger(__file__)
@@ -64,6 +65,7 @@ def driving_smarts_competition_v0_env(
     visdom: bool = False,
     sumo_headless: bool = True,
     envision_record_data_replay_path: Optional[str] = None,
+    observation_options: Union[ObservationOptions, str] = ObservationOptions.default,
 ):
     """An environment with a mission to be completed by a single or multiple ego agents.
 
@@ -188,19 +190,22 @@ def driving_smarts_competition_v0_env(
             ),
         )
 
+    observation_options = ObservationOptions[observation_options]
     env = HiWayEnvV1(
         scenarios=[env_specs["scenario"]],
         agent_interfaces=agent_interfaces,
         sim_name="Driving_SMARTS_v0",
         headless=headless,
         visdom=visdom,
+        fixed_timestep_sec=0.1,
         seed=seed,
         sumo_options=SumoOptions(headless=sumo_headless),
         visualization_client_builder=visualization_client_builder,
+        observation_options=observation_options,
     )
     env.action_space = env_action_space
     if ActionSpaceType[action_space] == ActionSpaceType.TargetPose:
-        env = _LimitTargetPose(env)
+        env = _LimitTargetPose(env, observation_options=observation_options)
     return env
 
 
@@ -395,13 +400,18 @@ def resolve_agent_interface(
 class _LimitTargetPose(gym.Wrapper):
     """Uses previous observation to limit the next TargetPose action range."""
 
-    def __init__(self, env: gym.Env):
+    def __init__(
+        self,
+        env: gym.Env,
+        observation_options: ObservationOptions = ObservationOptions.default,
+    ):
         """
         Args:
             env (gym.Env): Environment to be wrapped.
         """
         super().__init__(env)
         self._prev_obs: Dict[str, Dict[str, Any]]
+        self._observation_options = observation_options
 
     def step(
         self, action: Dict[str, np.ndarray]
@@ -451,12 +461,19 @@ def reset(self, **kwargs):
 
     def _store(self, obs: Dict[str, Any]) -> Dict[str, Dict[str, Any]]:
         filtered_obs: Dict[str, Dict[str, Any]] = {}
+
+        if self._observation_options == ObservationOptions.unformatted:
+            filter_func = lambda agent_obs: copy.deepcopy(
+                agent_obs.ego_vehicle_state.position[:2]
+            )
+        else:
+            filter_func = lambda agent_obs: copy.deepcopy(
+                agent_obs["ego_vehicle_state"]["position"][:2]
+            )
+
         for agent_name, agent_obs in obs.items():
-            filtered_obs[agent_name] = {
-                "position": copy.deepcopy(
-                    agent_obs["ego_vehicle_state"]["position"][:2]
-                )
-            }
+            filtered_obs[agent_name] = {"position": filter_func(agent_obs)}
+
         return filtered_obs
 
     def _limit(