schema_version: mmw_rf_compiler_main_training_policy_v1 base_config: configs/data_generation/data_gen_v1.0/config_smoke.yaml towns: - Town01 - Town02 - Town03 - Town15 states: dynamic_speeds_kmh: [10, 20, 40, 60] sampling: target_per_town_state: 5 # Coordinate alignment is now fixed and qualification runs show that empty # CIR is exceptional. Generate exactly the target count in the first pass. # A completed first pass may have a formal shortfall; later rounds generate # fresh CARLA geometry for exactly the remaining per-profile deficits. candidate_pool_factor: 1.0 # Close the CARLA geometry pool before any Sionna ray tracing. Full 2 kHz # oriented-footprint overlap QA rejects unsafe trajectories, and fresh CARLA # candidates are generated until every Town/state bucket has five safe runs. carla_trajectory_closure: enabled: true candidate_factor_per_deficit: 2.0 max_rounds: 8 max_no_progress_rounds: 2 supplement: enabled: true candidate_factor_per_deficit: 1.0 max_rounds: 5 max_no_progress_rounds: 2 main_training: profile_weights: core: 0.4 multipath: 0.2 scattering_mild: 0.2 scattering_medium: 0.2 vehicles: cav_blueprint: vehicle.tesla.model3 cav_rf_lod_index: 1 traffic_speed_min_kmh: 8.0 traffic_speed_max_kmh: 30.0 background_count_min: 8 background_count_max: 16 # Keep traffic inside the same local RF/camera operating region as the CAV. # This is especially important for Town15, whose global spawn-point catalogue # extends far beyond the compiled central ROI. background_rsu_distance_max_m: 120.0 # Reserve room for up to 1 s of background motion so a vehicle selected at # the boundary cannot leave the formal RF/camera range during capture. background_rsu_distance_margin_m: 10.0 # Some maps expose only sparse, map-wide native spawn points. Supplement # them from the OpenDRIVE driving-lane graph so every current or future Town # can populate the same local RSU region without Town-specific coordinates. background_spawn_waypoint_spacing_m: 8.0 background_count_strict: true # Probe-spawn above the sampled road surface, read the blueprint-specific # bounding box, then place the vehicle at exact ground contact. This avoids # collision rejection when a Town's waypoint Z is below the road mesh. vehicle_spawn_probe_lift_m: 3.0 # Prescribed trajectories do not use online collision avoidance because it # would make geometry depend on the simulation tick rate. Instead, derive a # route-independent safe spawn distance from the largest compiled vehicle # footprint, both actors' speed bounds, and the capture horizon. The extra # metre covers mesh/collision-envelope and floating-point differences. trajectory_spawn_clearance_margin_m: 1.0 # Town02's OpenDRIVE waypoint height is below the rendered/collision road # surface. Use a pre-spawn downward ray to establish the actual road height, # then preserve that surface-to-waypoint offset along each kinematic route. ground_alignment_mode: ground_projection ground_clearance_tolerance_m: 0.03 random_seed: 55 # Keep the dynamic shape topology stable inside a Town/state bucket so the # grouped Sionna resident-scene path can actually batch samples. fleet_grouping_scope: town_state acceleration: carla: workers: 2 gpu_devices: "0,1" shard_strategy: round_robin sionna: offline_qa: enabled: true profiles: [core, multipath, scattering_mild, scattering_medium] strict_profiles: [core, multipath] samples_per_bucket: 1 frames_per_sample: 32 statistical_profile_policy_file: configs/data_generation/data_gen_v1.0/config.yaml geometry_rate_hz: 2000 sample_duration_s: 1.0 warmup_s: 0.3 birdview_enabled: true birdview: enabled: true rate_hz: 20 height_m: 130.0 image_size_x: 1280 image_size_y: 1280 fov: 90.0 yaw_deg: 0.0 sensors: rsu_height_m: 5.0 camera_rate_hz: 20 lidar_rate_hz: 20 radar_rate_hz: 20 gnss_rate_hz: 10 imu_rate_hz: 100 camera_fov_deg: 110 cav_rgb_views: 4 cav_depth_views: 4 rsu_rgb_views: 4 rsu_depth_views: 4 rsu_camera_view_modes: [east, north, west, south] cav_lidar_vertical_fov_deg: [-10, 90] rsu_lidar_vertical_fov_deg: [-50, 50] rsu_radar_horizontal_fov_deg: 120 rsu_radar_vertical_fov_deg: 30 rsu_radar_yaw_policy: face_road # Probe the compiler-owned static PLYs before CARLA starts. The first offset # preserves the generated roadside; negative offsets select the opposite side # along the same road-normal axis. A scenario is not recorded unless one of # these deterministic positions has open sky, wall clearance, and plausible # ground below the 5 m device point. rsu_static_clearance: enabled: true lateral_offset_candidates_m: [8, -8, 10, -10, 6, -6, 12, -12, 4, -4] minimum_horizontal_clearance_m: 3.0 minimum_overhead_clearance_m: 4.0 horizontal_ray_count: 16 ground_probe_distance_m: 8.0 minimum_ground_distance_m: 4.0 maximum_ground_distance_m: 6.5 # Some compiler assets (currently Town15 at this road point) omit the road # triangle. A missing downward hit is audited as an asset-coverage warning; # any represented surface that is too close below the device still fails. require_ground_hit: false radio: center_frequency_hz: 3700000000 bandwidth_hz: 15360000 subcarrier_spacing_hz: 30000 num_subcarriers: 512 tx_antennas: 32 rx_antennas: 2 sionna: core_num_samples: 1000000 core_max_depth: 2 multipath_num_samples: 2000000 multipath_max_depth: 3 scattering_num_samples: 2000000 scattering_max_depth: 3 # Smoke uses the same solver budget as formal core generation. A reduced # ray count can turn a valid weak link into an artificial empty CIR. smoke_num_samples: 1000000 workers: 2 worker_policy_mode: adaptive worker_policy_file: configs/data_generation/data_gen_v2.0/rf_compiler_sionna_worker_policy_2x4090.json worker_policy_fingerprint: strict # Town15's current compiler asset is a 300 m x 300 m central ROI rather than a # full-map export. Endpoints are checked in Sionna's (x, -y, z) frame before RT. rf_support: Town15: min_rf_xyz_m: [64.566635, -348.369385, 80.0] max_rf_xyz_m: [364.566635, -48.369385, 200.0] endpoint_margin_m: 5.0 smoke: duration_s: 0.1