Environments

Environments#

This section describes the different categories of environments available in URBAN-SIM.

Environments for RL Training#

These environments are designed for training agents using reinforcement learning. They follow the standard Gymnasium interface and support batched simulation.

Base class: AbstractRLEnv in urbansim/envs/abstract_rl_env.py
Inheritance: Subclass of Isaac Lab’s ManagerBasedRLEnv
Configuration: Bound to a subclass of UrbanSceneCfg which defines and spawns assets, maps, agents, and sensors, etc.

To spawn such environment for training, we need to define the class as provided in urbansim/learning/RL/train.py:

@configclass
class EnvCfg(ManagerBasedRLEnvCfg):
    """Configuration for the locomotion velocity-tracking environment."""

    # Scene settings
    if pg_config is None:
        scene = scene_cfg(num_envs=env_config['Training']['num_envs'],
                          env_spacing=env_config['Omniverse']['env_spacing'],)
    else:
        scene = scene_cfg(num_envs=env_config['Training']['num_envs'],
                        env_spacing=env_config['Omniverse']['env_spacing'],
                        pg_config=pg_config,
                        scenario_generation_method=env_config['Omniverse'].get('scenario_generation_method', None),)
    # Basic settings
    viewer = ViewerCfg()
    observations = observation_cfg()
    actions = action_cfg()
    commands = command_cfg()
    # MDP settings
    rewards = reward_cfg()
    terminations = termination_cfg()
    events = event_cfg()
    curriculum = curriculum_cfg()

    def __post_init__(self):
        """Post-initialization to set up the environment."""
        super().__post_init__()
        # Additional setup can be done here if needed
        ...

and register it in gymnasium:

import gymnasium as gym
gym.register(
    id=f"URBANSIM-{task_name}-{robot_name}-{setting_name}",
    entry_point="urbansim.envs:AbstractRLEnv",
    disable_env_checker=True,
    kwargs={
        "env_cfg_entry_point": EnvCfg,
        "rsl_rl_cfg_entry_point":  f"urbansim.primitives.robot.{robot_name}:PPORunnerCfg",
        "rl_games_cfg_entry_point": f"configs/rl_configs/{task_name}/{robot_name}/{setting_name}_train.yaml",
    },
)

Environment Generation Pipeline#

URBAN-SIM supports two distinct pipelines for scenario generation, each suited to different use cases in training and evaluation.

Environments with Random Object Placement#

This lightweight pipeline places static props, buildings, and pedestrians at randomly sampled positions within a predefined region. It is ideal for large-scale training and introduces high scene diversity. Key characteristics include: - Fast initialization - Parameterized randomness (e.g., density, seed) - Low spatial structure, suitable for robust policy training

Run with:

python urbansim/envs/separate_envs/random_env.py --enable_cameras --num_envs 16 --use_async

Environments with Procedural Generation#

The PG pipeline creates structured scenes using rule-based or programmatic layouts. It supports curriculum learning and consistent benchmarking, thanks to repeatable spatial patterns. It is especially useful for: - Generalization tasks - Scene logic control - Progressive difficulty setups