Merged commit includes the following changes:

618016342 by Waymax Team: Internal change 618002526 by Waymax Team: Internal change 616933148 by Waymax Team: Internal change 616919769 by Waymax Team: Internal change 616835678 by Waymax Team: Internal change 609048090 by Waymax Team: Internal change 607734922 by Waymax Team: Internal change 607728802 by Waymax Team: Internal change 605672812 by Waymax Team: Internal change 602879589 by Waymax Team: Internal change 602828245 by Waymax Team: Internal change 602505506 by Waymax Team: Internal change 600572230 by Waymax Team: Internal change 600551386 by Waymax Team: Internal change 599209048 by Waymax Team: Internal change 586355223 by Waymax Team: Internal change 582747594 by Waymax Team: Internal change 578688603 by Waymax Team: Internal change 578278073 by Waymax Team: Internal change PiperOrigin-RevId: 618016342
waymo-research · Mar 22, 2024 · 20e3b11 · 20e3b11
1 parent baef8c3
commit 20e3b11
Show file tree

Hide file tree

Showing 28 changed files with 576 additions and 364 deletions.
diff --git a/LICENSE b/LICENSE
@@ -84,7 +84,14 @@ https://github.com/waymo-research/waymax/blob/main/LICENSE,
 and your access and use of the Waymx Licensed Materials are governed by the
 terms and conditions contained therein.
 
-@inproceedings{waymax, title={Waymax: An Accelerated, Data-Driven Simulator for Large-Scale Autonomous Driving Research}, author={Cole Gulino and Justin Fu and Wenjie Luo and George Tucker and Eli Bronstein and Yiren Lu and Jean Harb and Xinlei Pan and Yan Wang and Xiangyu Chen and John D. Co-Reyes and Rishabh Agarwal and Rebecca Roelofs and Yao Lu and Nico Montali and Paul Mougin and Zoey Yang and Brandyn White and Aleksandra Faust, and Rowan McAllister and Dragomir Anguelov and Benjamin Sapp}, booktitle={Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks},year={2023}}
+@inproceedings{waymax, title={Waymax: An Accelerated, Data-Driven Simulator for
+Large-Scale Autonomous Driving Research}, author={Cole Gulino and Justin Fu and
+Wenjie Luo and George Tucker and Eli Bronstein and Yiren Lu and Jean Harb and
+Xinlei Pan and Yan Wang and Xiangyu Chen and John D. Co-Reyes and Rishabh
+Agarwal and Rebecca Roelofs and Yao Lu and Nico Montali and Paul Mougin and
+Zoey Yang and Brandyn White and Aleksandra Faust, and Rowan McAllister and
+Dragomir Anguelov and Benjamin Sapp}, booktitle={Proceedings of the Neural
+Information Processing Systems Track on Datasets and Benchmarks},year={2023}}
 
 ii. In any license granting or any agreement governing use or access to Your
 Derivative IP, You must, and must require recipients of Your Derivative IP to

diff --git a/README.md b/README.md
@@ -18,7 +18,9 @@ distill behavior research into its simplest form.
 
 As all components are entirely written in JAX, Waymax is easily distributed and
 deployed on hardware accelerators, such as GPUs and
-[TPUs](https://cloud.google.com/tpu).
+[TPUs](https://cloud.google.com/tpu). Waymax is provided free of charge under
+the terms of the [Waymax License Agreement for Non-Commercial Use](https://github.com/waymo-research/waymax/blob/main/LICENSE).
+
 
 ## Installation
 
@@ -36,7 +38,7 @@ instructions on how to setup JAX with GPU/CUDA support if needed.
 
 Waymax is designed to work with the Waymo Open Motion dataset out of the box.
 
-A simple way to configure access is the following:
+A simple way to configure access via command line is the following:
 
 1.  Apply for [Waymo Open Dataset](https://waymo.com/open) access.
 
@@ -46,6 +48,13 @@ A simple way to configure access is the following:
 
 4.  Run `gcloud auth application-default login`.
 
+If you are using [colab](https://colab.google), run the following inside of the colab after registering in step 1:
+
+```python
+from google.colab import auth
+auth.authenticate_user()
+```
+
 Please reference
 [TF Datasets](https://www.tensorflow.org/datasets/gcs#authentication) for
 alternative methods to authentication.
@@ -151,3 +160,7 @@ Brandyn White and Aleksandra Faust, and Rowan McAllister and Dragomir Anguelov a
 booktitle={Proceedings of the Neural Information Processing Systems Track on Datasets and
 Benchmarks},year={2023}}
 ```
+
+## Contact
+
+Please email any questions to [[email protected]](mailto:[email protected]), or raise an issue on Github.
diff --git a/waymax/agents/actor_core.py b/waymax/agents/actor_core.py
@@ -14,12 +14,11 @@
 
 """Abstract definition of a Waymax actor for use at inference-time."""
 import abc
-from typing import Callable, TypeVar, Sequence
+from typing import Callable, Sequence, TypeVar
 
 import chex
 import jax
 import jax.numpy as jnp
-
 from waymax import datatypes
 
 # This is the internal state for whatever the agent needs to keep as its state.
@@ -28,6 +27,7 @@
 # This is the dictionary of parameters passed into the model which represents
 # the parameters to run the network.
 Params = datatypes.PyTree
+Action = datatypes.PyTree
 
 
 @chex.dataclass(frozen=True)
@@ -45,7 +45,7 @@ class WaymaxActorOutput:
   """
 
   actor_state: ActorState
-  action: datatypes.Action
+  action: Action
   is_controlled: jax.Array
 
   def validate(self):

diff --git a/waymax/agents/constant_speed.py b/waymax/agents/constant_speed.py
@@ -108,7 +108,7 @@ def __init__(self, speed: float = 0.0):
     Args:
       speed: Speed in m/s to set as the speed for all agents.
     """
-    super().__init__()
+    super().__init__(invalidate_on_end=True)
     self._speed = speed
 
   def update_speed(

diff --git a/waymax/agents/waypoint_following_agent.py b/waymax/agents/waypoint_following_agent.py
@@ -60,8 +60,10 @@ def __init__(
       is_controlled_func: Optional[
           Callable[[datatypes.SimulatorState], jax.Array]
       ] = None,
+      invalidate_on_end: bool = False,
   ):
     super().__init__(is_controlled_func=is_controlled_func)
+    self.invalidate_on_end = invalidate_on_end
 
   def update_trajectory(
       self, state: datatypes.SimulatorState
@@ -129,30 +131,37 @@ def _get_next_trajectory_by_projection(
     next_xy, next_yaw, reached_last_waypoint = _project_to_a_trajectory(
         jnp.stack([next_x, next_y], axis=-1),
         log_traj,
-        extrapolate_traj=False,
+        extrapolate_traj=not self.invalidate_on_end,
     )
 
     # Freeze the speed for agents that have reached the last waypoint to
     # prevent drift.
+    if self.invalidate_on_end:
+      default_x_vel = jnp.zeros_like(cur_sim_traj.vel_x)
+      default_y_vel = jnp.zeros_like(cur_sim_traj.vel_y)
+    else:
+      default_x_vel = cur_sim_traj.vel_x
+      default_y_vel = cur_sim_traj.vel_y
     new_vel_x = jnp.where(
         reached_last_waypoint,
-        cur_sim_traj.vel_x,
+        default_x_vel,
         new_speed * jnp.cos(cur_sim_traj.yaw),
     )
     new_vel_y = jnp.where(
         reached_last_waypoint,
-        cur_sim_traj.vel_y,
+        default_y_vel,
         new_speed * jnp.sin(cur_sim_traj.yaw),
     )
 
     # Invalidate moving objects that have reached their final waypoint.
     # This is to avoid invalidating parked cars. Use a threshold velocity
     # since some sim agents will tell the parked cars to move forward since
     # nothing is in front (e.g. IDM).
-    moving_after_last_waypoint = reached_last_waypoint & (
-        new_speed > _STATIC_SPEED_THRESHOLD
-    )
-    valid = valid & ~moving_after_last_waypoint
+    if self.invalidate_on_end:
+      moving_after_last_waypoint = reached_last_waypoint & (
+          new_speed > _STATIC_SPEED_THRESHOLD
+      )
+      valid = valid & ~moving_after_last_waypoint
 
     next_traj = cur_sim_traj.replace(
         x=next_xy[..., 0],
@@ -204,23 +213,28 @@ def __init__(
       min_spacing: float = 2.0,
       safe_time_headway: float = 2.0,
       max_accel: float = 2.0,
-      max_deccel: float = 4.0,
+      max_decel: float = 4.0,
       delta: float = 4.0,
       max_lookahead: int = 10,
+      lookahead_from_current_position: bool = True,
       additional_lookahead_points: int = 10,
       additional_lookahead_distance: float = 10.0,
+      invalidate_on_end: bool = False,
   ):
-    super().__init__(is_controlled_func=is_controlled_func)
+    super().__init__(
+        is_controlled_func=is_controlled_func,
+        invalidate_on_end=invalidate_on_end,
+    )
     self.desired_vel = desired_vel
     self.min_spacing_s0 = min_spacing
     self.safe_time_headway = safe_time_headway
     self.max_accel = max_accel
-    self.max_deccel = max_deccel
+    self.max_decel = max_decel
     self.delta = delta
     self.max_lookahead = max_lookahead
+    self.lookahead_from_current_position = lookahead_from_current_position
     self.additional_lookahead_distance = additional_lookahead_distance
     self.additional_headway_points = additional_lookahead_points
-    self.total_lookahead = max_lookahead + additional_lookahead_points
 
   def update_speed(
       self, state: datatypes.SimulatorState, dt: float = _DEFAULT_TIME_DELTA
@@ -287,9 +301,18 @@ def _get_accel(
     log_waypoints.validate()
     obj_curr_traj.validate()
     # 1. Find the closest waypoint and slice the future from that waypoint.
-    traj = _find_reference_traj_from_log_traj(
-        cur_position, log_waypoints, self.max_lookahead
-    )
+    if self.lookahead_from_current_position:
+      traj = _find_reference_traj_from_log_traj(cur_position, obj_curr_traj, 1)
+      chex.assert_shape(traj.xyz, prefix_shape + (num_obj, 1, 3))
+      total_lookahead = 1 + self.additional_headway_points
+    else:
+      traj = _find_reference_traj_from_log_traj(
+          cur_position, log_waypoints, self.max_lookahead
+      )
+      chex.assert_shape(
+          traj.xyz, prefix_shape + (num_obj, self.max_lookahead, 3)
+      )
+      total_lookahead = self.max_lookahead + self.additional_headway_points
 
 
     if self.additional_headway_points > 0:
@@ -303,7 +326,7 @@ def _get_accel(
     # max_lookahead) between traj (..., num_objects, max_lookahead) and
     # obj_curr_traj (..., num_objects, 1). Make common shape for bboxes:
     # (..., num_objects, num_objects, max_lookahead, 5).
-    broadcast_shape = prefix_shape + (num_obj, num_obj, self.total_lookahead, 5)
+    broadcast_shape = prefix_shape + (num_obj, num_obj, total_lookahead, 5)
     traj_5dof = traj.stack_fields(['x', 'y', 'length', 'width', 'yaw'])
     traj_bbox = jnp.broadcast_to(
         jnp.expand_dims(traj_5dof, axis=-3), broadcast_shape
@@ -356,7 +379,7 @@ def _get_accel(
         cur_speed * self.safe_time_headway
         + cur_speed
         * (cur_speed - lead_vel)
-        / (2 * jnp.sqrt(self.max_accel * self.max_deccel)),
+        / (2 * jnp.sqrt(self.max_accel * self.max_decel)),
     )
     # Set 0 for free-road behaviour.
     s_star = jnp.where(
@@ -521,21 +544,21 @@ def project_point_to_traj(
     src_yaw = traj.yaw[idx]
     src_dir = jnp.stack([jnp.cos(src_yaw), jnp.sin(src_yaw)], axis=-1)
 
+    last_valid_idx = jnp.where(traj.valid, jnp.arange(traj.shape[0]), 0)
+    last_valid_idx = jnp.argmax(last_valid_idx, axis=-1)
+    last_point = traj.xy[last_valid_idx, :]
+    reached_last_point = (
+        jnp.linalg.norm(last_point - src_xy, axis=-1)
+        < _REACHED_END_OF_TRAJECTORY_THRESHOLD
+    )
+    # Secondary detection: If a vehicle strays too far from the traj,
+    # also mark it as reaching the end.
+    reached_last_point = jnp.logical_or(
+        reached_last_point, dist[idx] > _DISTANCE_TO_REF_THRESHOLD
+    )
+
     # Prevent points from extrapolating beyond traj.
-    reached_last_point = jnp.zeros_like(idx, dtype=jnp.bool_)
     if not extrapolate_traj:
-      last_valid_idx = jnp.where(traj.valid, jnp.arange(traj.shape[0]), 0)
-      last_valid_idx = jnp.argmax(last_valid_idx, axis=-1)
-      last_point = traj.xy[last_valid_idx, :]
-      reached_last_point = (
-          jnp.linalg.norm(last_point - src_xy, axis=-1)
-          < _REACHED_END_OF_TRAJECTORY_THRESHOLD
-      )
-      # Secondary detection: If a vehicle strays too far from the traj,
-      # also mark it as reaching the end.
-      reached_last_point = jnp.logical_or(
-          reached_last_point, dist[idx] > _DISTANCE_TO_REF_THRESHOLD
-      )
       src_dir = jnp.where(reached_last_point, jnp.zeros_like(src_dir), src_dir)
 
     # Shape: (2).

diff --git a/waymax/agents/waypoint_following_agent_test.py b/waymax/agents/waypoint_following_agent_test.py
@@ -285,17 +285,19 @@ def test_decelerates_near_collision(self):
     cur_speed = jnp.array([10.0, 10.0])
     cur_position = jax.tree_util.tree_map(lambda x: x[..., :1], objects)
     max_accel = 1.13
-    max_deccel = 1.78
+    max_decel = 1.78
     delta = 4.0
     desired_vel = 30.0
     result = waypoint_following_agent.IDMRoutePolicy(
         max_accel=max_accel,
-        max_deccel=max_deccel,
+        max_decel=max_decel,
         desired_vel=desired_vel,
         min_spacing=1.0,
         safe_time_headway=1.0,
         max_lookahead=6,
         delta=delta,
+        lookahead_from_current_position=False,
+        invalidate_on_end=True,
     )._get_accel(objects, objects.xyz[:, 0, :], cur_speed, cur_position)
     # First agent should yield to second agent.
     # Second agent for free-road behavior.
@@ -313,7 +315,7 @@ def test_free_road_behavior(self, max_accel, cur_speed, desired_speed):
     cur_position = jax.tree_util.tree_map(lambda x: x[..., :1], waypoints)
     result = waypoint_following_agent.IDMRoutePolicy(
         max_accel=max_accel,
-        max_deccel=max_accel,
+        max_decel=max_accel,
         desired_vel=desired_speed,
         max_lookahead=6,
         delta=delta,

diff --git a/waymax/config.py b/waymax/config.py
@@ -120,32 +120,12 @@ class MetricsConfig:
   """Config for the built-in Waymax Metrics functions.
 
   Attributes:
-    run_log_divergence: Whether log_divergence metric will be computed in the
-      `step` function.
-    run_overlap: Whether overlap metric will be computed in the `step` function.
-    run_offroad: Whether offroad metric will be computed in the `step` function.
-    run_sdc_wrongway: Whether wrong-way metric will be computed for SDC in the
-      `step` function. Note this is only for single-agent env currently since
-      there is no route for sim-agents in data.
-    run_sdc_progression: Whether progression metric will be computed for SDC in
-      the `step` function. Note this is only for single-agent env currently
-      since there is no route for sim-agents in data.
-    run_sdc_off_route: Whether the off-route metric will be computed for SDC in
-      the `step` function. Note this is only for single-agent env currently
-      since there is no route for sim-agents in data.
-    run_sdc_kinematic_infeasibility: Whether the kinematics infeasibility metric
-      will be computed for SDC in the `step` function. Note this is only for
-      single-agent env currently since other agents may have different dynamics
-      and cannot be evaluated using the current kinematics infeasibility metrics
+    metrics_to_run: A list of metric names to run. Available metrics are:
+      log_divergence, overlap, offroad, sdc_wrongway, sdc_off_route,
+      sdc_progression, kinematic_infeasibility. Additional custom metrics can be
+      registered with `metric_factory.register_metric`.
   """
-
-  run_log_divergence: bool = True
-  run_overlap: bool = True
-  run_offroad: bool = True
-  run_sdc_wrongway: bool = False
-  run_sdc_progression: bool = False
-  run_sdc_off_route: bool = False
-  run_sdc_kinematic_infeasibility: bool = False
+  metrics_to_run: tuple[str, ...] = ('log_divergence', 'overlap', 'offroad')
 
 
 @dataclasses.dataclass(frozen=True)
@@ -154,11 +134,7 @@ class LinearCombinationRewardConfig:
 
   Attributes:
     rewards: Dictionary of metric names to floats indicating the weight of each
-      metric to create a reward of a linear combination. Valid metric names are
-      taken from the MetricConfig and removing 'run_'. For example, to create a
-      reward using the progression metric, the name would have to be
-      'sdc_progression', since 'run_sdc_progression' is used in the config
-      above.
+      metric to create a reward of a linear combination.
   """
 
   rewards: dict[str, float]
@@ -263,9 +239,9 @@ class WaymaxConfig:
 
   def __post_init__(self):
     if not self.data_config.include_sdc_paths and (
-        self.env_config.metrics.run_sdc_wrongway
-        | self.env_config.metrics.run_sdc_progression
-        | self.env_config.metrics.run_sdc_off_route
+        ('sdc_wrongway' in self.env_config.metrics.metrics_to_run)
+        | ('sdc_progression' in self.env_config.metrics.metrics_to_run)
+        | ('sdc_off_route' in self.env_config.metrics.metrics_to_run)
     ):
       raise ValueError(
           'Need to set data_config.include_sdc_paths True in  '