Scene Configuration

Scene Configuration#

URBAN-SIM supports multiple pipelines for generating simulation scenes with varying complexity, spatial structure, and randomness. Each environment binds a scene configuration (SceneCfg) that determines what types of scenes are loaded, how they are arranged, and how they are reused across multiple environments.

Scene Binding#

Scene configuration is specified in the environment config via:

@configclass
class EnvCfg(ManagerBasedRLEnvCfg):
    ...
    scene = SceneCfg(...)

The scene class determines the number of environments, their spatial arrangement, and the scenario generation method used.

Scenario Generation Pipelines#

URBAN-SIM includes two main pipelines for generating scenes:

  1. Random Object Placement Fast and lightweight. Places obstacles, buildings, and pedestrians randomly within a defined area. - Used via: random_env.py - Suitable for large-scale training and pretraining

  2. Procedural Generation (PG) Structured scene generation based on layout rules and scenario graphs. - Used via: pg_env.py - Suitable for curriculum learning and benchmarking

Each pipeline supports:

  • Scene-level asset instantiation

  • Environment duplication (–num_envs)

  • Asynchronous or synchronous layout control

Example YAML Config#

In a YAML config, the scene section may look like:

Env:
 name: static
 map_region: 30
 num_objects: 25
Omniverse:
 headless: True
 simulation_dt: 0.005
 rendering_interval: 4
 decimation: 20
 episode_length_s: 30.0
 env_spacing: 31.0
 scenario_generation_method: limited async procedural generation

This will spawn environments with randomized object placement and layout variation.

Scene Customization#

Scene behavior can be further customized via:

  • Changing asset spawn positions or scale

  • Overriding pg_config in the environment config

  • Specifying deterministic vs stochastic layouts

You can also inject dynamic agents (e.g., pedestrians, moving vehicles) through scenario setup like

pg_config: dict = dict(
     type='dynamic', # [clean, static, dynamic]
     with_terrain=False,
     with_boundary=True,
     map_region=20,
     buffer_width=1,
     num_object=10,
     num_pedestrian=9,
     walkable_seed=0,
     non_walkable_seed=1,
     seed=0,
     unique_env_num=20,
     ped_forward_inteval=10,
     moving_max_t=80,
 )

Spatial Arrangement#

URBAN-SIM supports both:

  • Synchronous scenes: All environments share the same map and layout (ideal for evaluation)

  • Asynchronous scenes: Each environment is different (ideal for generalization training)

Each environment is placed on a grid, spaced by env_spacing, and receives its own camera, sensor, and physics handle.

More details on scene generation and configuration can be found in the file urbansim/scene/urban_scene.py.