General Configurations

General Configurations#

UrbanVerse uses a structured configuration system to define training environments for goal-directed navigation. Similar to the configuration style used in robotics simulation frameworks, UrbanVerse encapsulates all environment parameters—scenes, robot embodiment, observation and action spaces, rewards, and termination logic—under a unified configuration object.

The central configuration entry point is EnvCfg. This class specifies how an UrbanVerse navigation environment should be built, and is consumed by the high-level training API:

uv.navigation.rl.create_env(cfg)

This page introduces the roles of each configuration component, their hierarchy, and how they together define a full reinforcement learning environment.

EnvCfg: The Environment Schema#

EnvCfg is a lightweight configuration schema describing the full training environment:

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

cfg = EnvCfg(
    scenes=SceneCfg(scene_paths=my_scene_list),
    robot_type="coco_wheeled",
    observations=ObservationCfg(),
    actions=ActionCfg(),
    rewards=RewardCfg(),
    terminations=TerminationCfg(),
    curriculum=CurriculumCfg(),
)

Once initialized, an environment is created via:

env = uv.navigation.rl.create_env(cfg)

uv.navigation.rl.train(
    env=env,
    training_cfg=training_cfg,
    output_dir="outputs/ppo_run",
)

The remainder of this section explains each field in detail.

Scene Configuration#

UrbanVerse scenes are real-to-sim USD environments reconstructed by UrbanVerse-Gen or provided as part of UrbanVerse-160 / CraftBench. Scene configuration specifies which scenes to load and how they should be distributed across parallel simulation instances.

SceneCfg(
    scene_paths=[
        "/path/UrbanVerse-160/Asia_Japan_Tokyo_0001/scene_cousin_01/scene.usd",
        "/path/UrbanVerse-160/Africa_SouthAfrica_CapeTown_0002/scene_cousin_03/scene.usd",
        ...
    ],
    async_sim=True,
    env_spacing=3.0,
)

Key Parameters#

  • scene_paths A list of USD scenes. Each path may represent a different layout (city) and/or different digital cousin variants of the same layout.

  • async_sim

    • True: each environment loads a different scene in parallel (recommended) - ideal for generalization

    • False: all environments use the same layout but different cousin variants - ideal for debugging and controlled experiments

  • env_spacing The horizontal offset between parallel environments in the simulator.

UrbanVerse ensures that each training worker receives a valid scene and that scenes can be shuffled or rotated between episodes through the RL API.

Observation Configuration#

The UrbanVerse navigation task uses a compact sensor specification suitable for on-policy RL:

ObservationCfg(
    rgb_size=(135, 240),
    use_depth=False,
    include_goal_vector=True,
)

Fields#

  • rgb_size – resolution of the onboard camera

  • use_depth – whether a depth channel is included

  • include_goal_vector – the relative (dx, dy) distance to the goal is appended to the observation buffer

This aligns with the network architecture described in the UrbanVerse paper.

Action Configuration#

Actions define the control interface for the robot. UrbanVerse exposes a continuous velocity-command interface for coco_wheeled:

ActionCfg(
    action_dim=2,         # linear_x, angular_z
    linear_limit=1.0,
    angular_limit=1.0,
)

This configuration is passed into the environment and later used by the PPO actor network.

Reward Configuration#

Rewards in UrbanVerse reflect the design presented in the paper, balancing:

  • arrival reward

  • collision penalty

  • coarse & fine waypoint tracking

  • velocity alignment

These fields are bundled in RewardCfg:

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

This schema allows reproducible experiments across UrbanVerse scenes.

Termination Configuration#

UrbanVerse defines a minimal set of clear termination conditions:

  • success (goal reached)

  • collision

  • timeout (episode length exceeded)

  • out-of-bounds / leaving traversable regions

These are collected in:

TerminationCfg(
    max_episode_steps=300,
    enable_collision=True,
    enable_success=True,
    enable_timeout=True,
)

Curriculum Configuration#

An optional curriculum can be enabled to gradually increase training difficulty by:

  • expanding goal distance ranges

  • introducing lighting or cousin jitter

  • varying robot initialization poses

CurriculumCfg(
    enable_goal_distance_curriculum=True,
    enable_cousin_jitter=True,
)

Curriculum settings are optional but often accelerate early-stage learning.

Dynamic Initialization#

Robot spawn points are automatically sampled from drivable road regions of each scene:

robot_cfg.init_state.pos = sample_valid_spawn_point()
robot_cfg.init_state.yaw = sample_valid_heading()

This ensures valid initialization across diverse city layouts.

Summary#

  • EnvCfg defines the entire UrbanVerse training environment

  • It specifies scenes, robots, observations, actions, rewards, terminations, and optional curriculum

  • Configuration style mirrors established robotics-simulation documentation practices

  • The API remains simple:

    env = uv.navigation.rl.create_env(cfg) uv.navigation.rl.train(env, ...)

This structuring ensures clarity, reproducibility, and extensibility for robot learning research in UrbanVerse.