Configuration API

Configuration API#

The Configuration API provides classes for defining RL environment configurations in UrbanVerse.

Import#

from urbanverse.navigation.config import (
    EnvCfg, SceneCfg, ObservationCfg, ActionCfg,
    RewardCfg, TerminationCfg, CurriculumCfg, EventCfg
)

EnvCfg#

EnvCfg(
    scenes: SceneCfg,
    robot_type: str,
    observations: ObservationCfg,
    actions: ActionCfg,
    rewards: RewardCfg,
    terminations: TerminationCfg,
    curriculum: CurriculumCfg | None = None,
    events: EventCfg | None = None,
)

Central configuration class that defines the complete RL environment.

Parameters:

  • scenes (SceneCfg): Scene configuration (see below)

  • robot_type (str): Robot embodiment identifier

  • observations (ObservationCfg): Observation space configuration (see below)

  • actions (ActionCfg): Action space configuration (see below)

  • rewards (RewardCfg): Reward function configuration (see below)

  • terminations (TerminationCfg): Termination conditions configuration (see below)

  • curriculum (CurriculumCfg, optional): Curriculum learning configuration (see below)

  • events (EventCfg, optional): Event configuration (see below)

Example:

cfg = EnvCfg(
    scenes=SceneCfg(scene_paths=["/path/to/scene.usd"], async_sim=True),
    robot_type="coco_wheeled",
    observations=ObservationCfg(rgb_size=(135, 240)),
    actions=ActionCfg(),
    rewards=RewardCfg(),
    terminations=TerminationCfg(max_episode_steps=300),
    curriculum=CurriculumCfg(enable_goal_distance_curriculum=True),
)

SceneCfg#

SceneCfg(
    scene_paths: list[str],
    async_sim: bool = True,
    env_spacing: float = 3.0,
)

Configuration for scene loading and distribution.

Parameters:

  • scene_paths (list[str]): List of USD scene file paths

  • async_sim (bool, optional): Scene distribution mode. Default: True - True: Each environment loads a different scene (recommended for generalization) - False: All environments use the same layout but different cousin variants

  • env_spacing (float, optional): Horizontal spacing between parallel environments (meters). Default: 3.0

ObservationCfg#

ObservationCfg(
    rgb_size: tuple[int, int] = (135, 240),
    use_depth: bool = False,
    include_goal_vector: bool = True,
)

Configuration for observation space.

Parameters:

  • rgb_size (tuple[int, int], optional): RGB camera image resolution (height, width). Default: (135, 240)

  • use_depth (bool, optional): Whether to include depth channel. Default: False

  • include_goal_vector (bool, optional): Whether to include goal-relative position vector (dx, dy). Default: True

ActionCfg#

ActionCfg(
    action_dim: int = 2,
    linear_limit: float = 1.0,
    angular_limit: float = 1.0,
)

Configuration for action space.

Parameters:

  • action_dim (int, optional): Action dimension. Default: 2 (linear, angular for wheeled robots)

  • linear_limit (float, optional): Maximum linear velocity (m/s). Default: 1.0

  • angular_limit (float, optional): Maximum angular velocity (rad/s). Default: 1.0

Note: Action space is automatically adapted based on robot type. For legged/humanoid robots, joint-level controls are used.

RewardCfg#

RewardCfg(
    arrived_reward: float = 2000.0,
    collision_penalty: float = -200.0,
    tracking_fine_std: float = 1.0,
    tracking_fine_weight: float = 50.0,
    tracking_coarse_std: float = 5.0,
    tracking_coarse_weight: float = 10.0,
    velocity_weight: float = 10.0,
)

Configuration for reward function.

Parameters:

  • arrived_reward (float, optional): Reward for reaching goal. Default: 2000.0

  • collision_penalty (float, optional): Penalty for collisions. Default: -200.0

  • tracking_fine_std (float, optional): Standard deviation for fine waypoint tracking (meters). Default: 1.0

  • tracking_fine_weight (float, optional): Weight for fine tracking reward. Default: 50.0

  • tracking_coarse_std (float, optional): Standard deviation for coarse waypoint tracking (meters). Default: 5.0

  • tracking_coarse_weight (float, optional): Weight for coarse tracking reward. Default: 10.0

  • velocity_weight (float, optional): Weight for velocity alignment reward. Default: 10.0

TerminationCfg#

TerminationCfg(
    max_episode_steps: int = 300,
    enable_collision: bool = True,
    enable_success: bool = True,
    enable_timeout: bool = True,
    enable_out_of_bounds: bool = False,
)

Configuration for episode termination conditions.

Parameters:

  • max_episode_steps (int, optional): Maximum steps before timeout. Default: 300

  • enable_collision (bool, optional): Terminate on collision. Default: True

  • enable_success (bool, optional): Terminate on goal reached. Default: True

  • enable_timeout (bool, optional): Terminate on step limit. Default: True

  • enable_out_of_bounds (bool, optional): Terminate when leaving traversable regions. Default: False

CurriculumCfg#

CurriculumCfg(
    enable_goal_distance_curriculum: bool = False,
    enable_cousin_jitter: bool = False,
)

Configuration for curriculum learning.

Parameters:

  • enable_goal_distance_curriculum (bool, optional): Gradually increase goal distances. Default: False

  • enable_cousin_jitter (bool, optional): Gradually increase scene diversity. Default: False

EventCfg#

EventCfg(
    enable_robot_state_randomization: bool = False,
    enable_dynamic_agent_spawning: bool = False,
)

Configuration for environment events and randomization.

Parameters:

  • enable_robot_state_randomization (bool, optional): Randomize robot initial state. Default: False

  • enable_dynamic_agent_spawning (bool, optional): Spawn dynamic agents during episodes. Default: False

Note: Robot spawn points are automatically sampled from drivable regions. This configuration is for additional randomization.