Environment Configurations

Environment Configurations#

The EnvCfg class defines the configuration schema for URBAN-SIM training environments. It inherits from ManagerBasedRLEnvCfg in Isaac Lab and encapsulates settings for the simulation scene, observation/action space, rewards, curriculum, and termination logic.

@configclass
class EnvCfg(ManagerBasedRLEnvCfg):
    """
    Environment configuration schema for URBAN-SIM.
    Inherits from ManagerBasedRLEnvCfg.
    """

    # Scene configuration
    scene: SceneCfg = SceneCfg()

    # Simulation parameters
    viewer: ViewerCfg = ViewerCfg()
    observations: ObservationCfg = ObservationCfg()
    actions: ActionCfg = ActionCfg()
    commands: CommandCfg = CommandCfg()

    # MDP components
    rewards: RewardCfg = RewardCfg()
    terminations: TerminationCfg = TerminationCfg()
    events: EventCfg = EventCfg()
    curriculum: CurriculumCfg = CurriculumCfg()

    # Robot configuration
    robot_name: str = "COCO"  # Default robot type

Scene Configuration#

The scene field instantiates a SceneCfg-compatible class, determining how the environment is constructed.

  • If pg_config is not provided, the system defaults to random object placement:

    scene = scene_cfg(num_envs=..., env_spacing=...)

  • If pg_config is provided, a procedural scenario is built:

    scene = scene_cfg(
    num_envs=...,
    env_spacing=...,
    pg_config=pg_config,
    scenario_generation_method=...,
)

Key options:

  • num_envs: Number of parallel environments

  • env_spacing: Spacing between environments in simulation world

  • pg_config: Optional procedural generation config

  • scenario_generation_method: Overrides default random placement

Simulation Configuration#

  • viewer: Defines viewer resolution, camera mode, etc. via ViewerCfg

  • observations: Sensor and observation space definitions via observation_cfg

  • actions: Action space definition (e.g., continuous, discrete) via action_cfg

  • commands: External command signal structure via command_cfg

MDP Configuration#

  • rewards: Reward shaping logic via reward_cfg

  • terminations: Termination condition logic via termination_cfg

  • events: Optional triggerable in-sim events

  • curriculum: Curriculum learning parameters via curriculum_cfg

All these components are dynamically loaded based on the environment configuration, allowing for flexible and extensible environment setups.

Robot Configuration#

URBAN-SIM supports multiple robot embodiments, each with its own physical parameters, control interface, and integration strategy.

The robot selection is determined via the robot_name field in the environment configuration, and dynamically loads the corresponding config modules.

Supported Robots#

  1. COCO (wheeled base)

    • Config: COCO_CFG from urbansim.primitives.robot.coco

    • Action space: COCOVelocityActionsCfg

    • Environment modifier: COCONavModifyEnv

    • Default height: z = 0.4

  2. Unitree Go2 (quadruped robot)

    • Config: UNITREE_GO2_CFG from urbansim.primitives.robot.unitree_go2

    • Action space: GO2NavActionsCfg

    • Environment modifier: GO2NavModifyEnv

    • Default height: z = 0.3

  3. Unitree G1 (humanoid / bipedal)

    • Config: G1_MINIMAL_CFG from urbansim.primitives.robot.unitree_g1

    • Action space: G1NavActionsCfg

    • Environment modifier: G1NavModifyEnv

    • Default height: z = 0.74

Dynamic Initialization#

The configuration system selects robot-specific components based on name:

if robot_name.lower() == "unitree_go2":
    from urbansim.primitives.robot.unitree_go2 import UNITREE_GO2_CFG, GO2NavActionsCfg, GO2NavModifyEnv
    robot_cfg = UNITREE_GO2_CFG
    action_cfg = GO2NavActionsCfg
    modify_env_fn = GO2NavModifyEnv

# Set robot spawn position
robot_cfg.init_state.pos = env_config["Robot"].get("init_position", default_xyz)

Action Configuration#

Each robot defines its own action_cfg class, determining:

  • Control mode (e.g., velocity commands, joint torques)

  • Action dimension and limits

  • Mapping to simulation API

These configurations are injected into the full environment config (e.g., EnvCfg) to ensure proper wiring during instantiation.