# Network description of the OnePlanner model in this directory (oneplanner_planner.onnx and
# bevfusion_lidar_feature.onnx), read by autoware_oneplanner as ml_package_<model_name>.param.yaml.
# It overrides the package defaults for this network; host and deployment behaviour (workspace,
# rates, tolerances, paths) stays in the package's oneplanner.param.yaml. See the package's
# docs/model_contract.md for the meaning of every field.
/**:
  ros__parameters:
    # The input convention the model was trained with; the node refuses any other.
    input_convention: "oneplanner_v1"
    # Which sensor input providers the model needs, matched by tensor name.
    sensor_inputs: ["bev_feature"]
    enable_context_inputs: true
    # When the plan starts: at the newest odometry; the graph reads sensor_latency.
    planning_time: "planning_time"

    # Builder precision of the planner graph.
    precision: "fp32"

    context:
      ego_history_interval_seconds: 0.1
      # The graph does not read ego_curvature_bias.
      curvature_bias:
        enabled: false
        tau_s: 2.0
        min_speed_mps: 1.0
      # The graph does not read lane_signals / route_signals.
      signal_history:
        enabled: false
        steps: 7
        step_s: 0.5
      turn_indicators:
        # The graph carries this input for a stable signature only; perturbing it changes no
        # output. Bind a constant, subscribe to nothing.
        enabled: false

    bev_feature:
      history_tensor: "bev_feature_history"
      frames: 3
      interval_seconds: 0.5
      bev_half_extent_m: 122.4
      extractor:
        precision: "fp16"
        feature_tensor: "bev_feature"
        max_points_per_voxel: 32  # from the voxels input shape
        use_intensity: true  # 5 point features
        voxels_num: [1, 128000, 256000]  # [min, opt, max]
        point_cloud_range: [-122.4, -122.4, -3.0, 122.4, 122.4, 5.0]
        voxel_size: [0.17, 0.17, 0.2]
      detection:
        # The frozen BEVFusion detection head the extractor graph carries. It emits undecoded
        # proposals; the node decodes and suppresses them with autoware_bevfusion's
        # postprocessing and publishes ~/output/detected_objects.
        enabled: true
        # One Autoware class name per head class, in the head's label order. A head class
        # with no ObjectClassification label is written as UNKNOWN.
        # head classes: car, truck, bus, bicycle, pedestrian, traffic_cone, barrier
        class_names: ["CAR", "TRUCK", "BUS", "BICYCLE", "PEDESTRIAN", "UNKNOWN", "UNKNOWN"]
        num_proposals: 500
        out_size_factor: 8
        # Score thresholds per (distance band, class), calibrated for this head and applied as
        # autoware_bevfusion applies detection_score_thresholds: the box centre's radial
        # distance picks the first band whose upper limit exceeds it, and a box beyond the last
        # band is dropped.
        distance_bin_upper_limits: [50.0, 90.0, 121.0, 200.0]
        # distance-major: one row per band, one column per head class
        # CAR       TRUCK     BUS       BICYCLE   PEDESTRIAN  TRAFFIC_CONE  BARRIER
        score_thresholds: [
          0.194733, 0.167118, 0.109791, 0.245172, 0.150065, 0.12842, 0.043799,  # [0, 50) m
          0.221063, 0.18848, 0.168961, 0.160301, 0.152223, 0.15707, 0.1,  # [50, 90) m
          0.155292, 0.163172, 0.08902, 0.185649, 0.154218, 0.239001, 0.1,  # [90, 121) m
          0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1  # [121, 200) m
        ]

    postprocess:
      prediction_tensor: "trajectory"
      extra_trajectory_tensors: ["prior_trajectory"]
      # Logits of DISABLE, ENABLE_LEFT, ENABLE_RIGHT, published as ~/output/turn_indicators.
      turn_indicator:
        tensor: "turn_indicator_logits"
      horizon_seconds: 6.0  # 60 steps
      time_step: 0.1
      generator_name: "OnePlanner"