ParkingRobot/data_process/plot_new_controller_experiment.py

"""为新版控制器实验CSV生成包含轨迹、误差、速度和转角的六子图总图。"""

from __future__ import annotations

import argparse
from pathlib import Path

import matplotlib

matplotlib.use("Agg")
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd


SCRIPT_DIR = Path(__file__).resolve().parent
LATERAL_JUMP_THRESHOLD_METERS = 0.03
JUMP_INSET_CONTEXT_SAMPLES = 6
MAXIMUM_PLAUSIBLE_LINEAR_SPEED_METERS_PER_SECOND = 1.20
POSITION_JUMP_MARGIN_METERS = 0.03


def configure_matplotlib() -> None:
    """配置可显示中文和负号的Matplotlib字体。"""
    plt.rcParams["font.sans-serif"] = [
        "Microsoft YaHei",
        "SimHei",
        "Noto Sans CJK SC",
        "Arial Unicode MS",
        "DejaVu Sans",
    ]
    plt.rcParams["axes.unicode_minus"] = False


def numeric_column(
    frame: pd.DataFrame,
    name: str,
    default: float = np.nan,
) -> np.ndarray:
    """将CSV列安全转换为浮点数组，缺失列使用指定默认值。"""
    if name not in frame.columns:
        return np.full(len(frame), default, dtype=float)
    return pd.to_numeric(frame[name], errors="coerce").to_numpy(
        dtype=float,
        copy=True,
    )


def first_finite(values: np.ndarray, default: float) -> float:
    """读取数组中的第一个有限值。"""
    finite = values[np.isfinite(values)]
    return float(finite[0]) if finite.size else default


def first_text(frame: pd.DataFrame, name: str, default: str) -> str:
    """读取文本元数据列中的第一个非空值。"""
    if name not in frame.columns:
        return default
    values = frame[name].dropna().astype(str)
    values = values[values.str.strip() != ""]
    return values.iloc[0] if not values.empty else default


def text_column(frame: pd.DataFrame, name: str) -> np.ndarray:
    """读取用于诊断标注的原始文本列，缺失值转换为空字符串。"""
    if name not in frame.columns:
        return np.full(len(frame), "", dtype=object)
    return frame[name].fillna("").astype(str).to_numpy(
        dtype=object,
        copy=True,
    )


def fill_reference_series(
    values: np.ndarray,
    fallback: np.ndarray,
) -> np.ndarray:
    """前后填充后台采样得到的控制参考值，缺失时使用解析速度曲线。"""
    series = pd.Series(values, dtype=float)
    filled = series.ffill().bfill().to_numpy(
        dtype=float,
        copy=True,
    )
    missing = ~np.isfinite(filled)
    filled[missing] = fallback[missing]
    return filled


def planned_motion(
    time_seconds: np.ndarray,
    length_meters: float,
    cruise_speed_mps: float,
    acceleration_mps2: float,
    deceleration_mps2: float,
) -> tuple[np.ndarray, np.ndarray]:
    """计算从静止出发并在终点静止的梯形或三角形理想时间速度轨迹。"""
    acceleration_distance = (
        cruise_speed_mps**2 / (2.0 * acceleration_mps2)
    )
    deceleration_distance = (
        cruise_speed_mps**2 / (2.0 * deceleration_mps2)
    )

    if acceleration_distance + deceleration_distance <= length_meters:
        peak_speed = cruise_speed_mps
    else:
        peak_speed = np.sqrt(
            2.0
            * length_meters
            / (1.0 / acceleration_mps2 + 1.0 / deceleration_mps2)
        )
        acceleration_distance = (
            peak_speed**2 / (2.0 * acceleration_mps2)
        )
        deceleration_distance = (
            peak_speed**2 / (2.0 * deceleration_mps2)
        )

    acceleration_time = peak_speed / acceleration_mps2
    deceleration_time = peak_speed / deceleration_mps2
    cruise_distance = max(
        0.0,
        length_meters - acceleration_distance - deceleration_distance,
    )
    cruise_time = cruise_distance / peak_speed
    deceleration_start_time = acceleration_time + cruise_time
    finish_time = deceleration_start_time + deceleration_time

    progress = np.zeros_like(time_seconds, dtype=float)
    speed = np.zeros_like(time_seconds, dtype=float)

    accelerating = time_seconds <= acceleration_time
    progress[accelerating] = (
        0.5 * acceleration_mps2 * time_seconds[accelerating] ** 2
    )
    speed[accelerating] = acceleration_mps2 * time_seconds[accelerating]

    cruising = (
        (time_seconds > acceleration_time)
        & (time_seconds <= deceleration_start_time)
    )
    progress[cruising] = (
        acceleration_distance
        + peak_speed * (time_seconds[cruising] - acceleration_time)
    )
    speed[cruising] = peak_speed

    decelerating = (
        (time_seconds > deceleration_start_time)
        & (time_seconds <= finish_time)
    )
    remaining_time = finish_time - time_seconds[decelerating]
    progress[decelerating] = (
        length_meters
        - 0.5 * deceleration_mps2 * remaining_time**2
    )
    speed[decelerating] = deceleration_mps2 * remaining_time

    finished = time_seconds > finish_time
    progress[finished] = length_meters
    speed[finished] = 0.0
    return progress, speed


def load_experiment(csv_path: Path) -> dict[str, object]:
    """读取新版CSV并构造绘图所需的统一SI单位数据。"""
    frame = pd.read_csv(csv_path, encoding="utf-8-sig")
    if frame.empty:
        raise ValueError("CSV没有任何采样行。")

    time_seconds = numeric_column(frame, "ElapsedSeconds")
    valid_time = np.isfinite(time_seconds)
    frame = frame.loc[valid_time].reset_index(drop=True)
    time_seconds = time_seconds[valid_time]
    if time_seconds.size < 2:
        raise ValueError("CSV中的有效时间采样不足2帧。")
    time_seconds = time_seconds - time_seconds[0]

    state_x = numeric_column(frame, "StateXMeters")
    state_y = numeric_column(frame, "StateYMeters")
    has_processed = numeric_column(frame, "HasProcessedState", 0.0) > 0.5
    processed_valid = has_processed & np.isfinite(state_x) & np.isfinite(state_y)

    raw_x_meters = numeric_column(frame, "DetourX") / 1000.0
    raw_y_meters = numeric_column(frame, "DetourY") / 1000.0
    valid_raw_position = (
        np.isfinite(raw_x_meters) & np.isfinite(raw_y_meters)
    )
    detour_tick_raw = text_column(frame, "DetourTickRaw")
    detour_l_step = numeric_column(frame, "DetourLStep")
    actual_x = np.where(processed_valid, state_x, raw_x_meters)
    actual_y = np.where(processed_valid, state_y, raw_y_meters)
    valid_position = np.isfinite(actual_x) & np.isfinite(actual_y)
    if np.count_nonzero(valid_position) < 2:
        raise ValueError("CSV中没有足够的有效车辆位置。")

    start = np.array(
        [
            first_finite(numeric_column(frame, "ReferenceStartX"), np.nan),
            first_finite(numeric_column(frame, "ReferenceStartY"), np.nan),
        ],
        dtype=float,
    ) / 1000.0
    end = np.array(
        [
            first_finite(numeric_column(frame, "ReferenceEndX"), np.nan),
            first_finite(numeric_column(frame, "ReferenceEndY"), np.nan),
        ],
        dtype=float,
    ) / 1000.0
    if not np.all(np.isfinite(start)) or not np.all(np.isfinite(end)):
        raise ValueError("CSV缺少有效的参考起点或终点。")

    line = end - start
    length_meters = float(np.linalg.norm(line))
    if length_meters <= 1e-6:
        raise ValueError("参考直线长度必须大于0。")
    tangent = line / length_meters
    left_normal = np.array([-tangent[1], tangent[0]])
    displacement = np.column_stack([actual_x, actual_y]) - start
    # 与C# TrajectoryProjector保持一致：轨迹位于车辆左侧时为正。
    derived_lateral_error = -(displacement @ left_normal)
    recorded_lateral_error = numeric_column(
        frame,
        "ControlLateralErrorMeters",
    )
    has_control_reference = (
        numeric_column(frame, "HasControlReference", 0.0) > 0.5
    )
    recorded_lateral_valid = (
        has_control_reference & np.isfinite(recorded_lateral_error)
    )
    lateral_error = (
        np.where(recorded_lateral_valid, recorded_lateral_error, np.nan)
        if np.any(recorded_lateral_valid)
        else derived_lateral_error
    )

    state_yaw = numeric_column(frame, "StateYawRadians")
    raw_yaw = np.deg2rad(numeric_column(frame, "DetourTheta"))
    actual_yaw = np.where(
        has_processed & np.isfinite(state_yaw),
        state_yaw,
        raw_yaw,
    )
    reference_yaw = np.arctan2(tangent[1], tangent[0])
    derived_heading_error = np.arctan2(
        np.sin(reference_yaw - actual_yaw),
        np.cos(reference_yaw - actual_yaw),
    )
    recorded_heading_error = numeric_column(
        frame,
        "ControlHeadingErrorRadians",
    )
    recorded_heading_valid = (
        has_control_reference & np.isfinite(recorded_heading_error)
    )
    heading_error = (
        np.where(recorded_heading_valid, recorded_heading_error, np.nan)
        if np.any(recorded_heading_valid)
        else derived_heading_error
    )

    # 投影定义满足：参考点 = 车体位置 + 横向误差 × 参考航向左法向。
    # 因此无需假设轨迹类型，即可从有效控制周期还原车辆实际使用的参考轨迹。
    projected_reference_yaw = actual_yaw + heading_error
    reference_x = (
        actual_x - lateral_error * np.sin(projected_reference_yaw)
    )
    reference_y = (
        actual_y + lateral_error * np.cos(projected_reference_yaw)
    )
    valid_reference_position = (
        has_control_reference
        & np.isfinite(reference_x)
        & np.isfinite(reference_y)
    )

    # 同时检查Detour是否超出车辆物理运动边界，以及控制状态的横向误差
    # 是否发生离散突变。后者能覆盖状态层延迟接受持续定位偏移的情况。
    raw_delta_x = np.full(len(frame), np.nan, dtype=float)
    raw_delta_y = np.full(len(frame), np.nan, dtype=float)
    sample_delta_time = np.full(len(frame), np.nan, dtype=float)
    raw_delta_x[1:] = np.diff(raw_x_meters)
    raw_delta_y[1:] = np.diff(raw_y_meters)
    sample_delta_time[1:] = np.diff(time_seconds)
    detour_position_step = np.hypot(raw_delta_x, raw_delta_y)

    detour_tick_numeric = numeric_column(frame, "DetourTickRaw")
    detour_tick_delta_seconds = np.full(len(frame), np.nan, dtype=float)
    detour_tick_delta_seconds[1:] = (
        np.diff(detour_tick_numeric) / 10_000_000.0
    )
    source_delta_time = sample_delta_time.copy()
    valid_tick_delta = (
        np.isfinite(detour_tick_delta_seconds)
        & (detour_tick_delta_seconds > 0.0)
        & (detour_tick_delta_seconds <= 0.5)
    )
    source_delta_time[valid_tick_delta] = (
        detour_tick_delta_seconds[valid_tick_delta]
    )

    consecutive_raw_position_valid = np.zeros(len(frame), dtype=bool)
    consecutive_raw_position_valid[1:] = (
        valid_raw_position[1:] & valid_raw_position[:-1]
    )
    maximum_plausible_position_step = (
        MAXIMUM_PLAUSIBLE_LINEAR_SPEED_METERS_PER_SECOND
        * source_delta_time
        + POSITION_JUMP_MARGIN_METERS
    )
    raw_detour_jump = (
        consecutive_raw_position_valid
        & np.isfinite(detour_position_step)
        & np.isfinite(source_delta_time)
        & (source_delta_time > 0.0)
        & (source_delta_time <= 0.5)
        & (detour_position_step > maximum_plausible_position_step)
    )

    state_lateral_step = np.full(len(frame), np.nan, dtype=float)
    state_lateral_step[1:] = np.diff(lateral_error)
    state_lateral_jump = (
        np.isfinite(state_lateral_step)
        & np.isfinite(sample_delta_time)
        & (sample_delta_time > 0.0)
        & (sample_delta_time <= 0.5)
        & (np.abs(state_lateral_step) >= LATERAL_JUMP_THRESHOLD_METERS)
    )
    suspected_jump = raw_detour_jump | state_lateral_jump
    jump_indices = np.flatnonzero(suspected_jump)
    jump_magnitude = np.zeros(len(frame), dtype=float)
    jump_magnitude[raw_detour_jump] = detour_position_step[raw_detour_jump]
    jump_magnitude[state_lateral_jump] = np.maximum(
        jump_magnitude[state_lateral_jump],
        np.abs(state_lateral_step[state_lateral_jump]),
    )

    cruise_speed = first_finite(
        numeric_column(frame, "ReferenceSpeed"),
        0.30,
    )
    acceleration = first_finite(
        numeric_column(
            frame,
            "ReferenceAccelerationMetersPerSecondSquared",
        ),
        0.20,
    )
    deceleration = first_finite(
        numeric_column(
            frame,
            "ReferenceDecelerationMetersPerSecondSquared",
        ),
        0.20,
    )
    if acceleration <= 0.0:
        acceleration = 0.20
    if deceleration <= 0.0:
        deceleration = 0.20

    _, ideal_speed = planned_motion(
        time_seconds,
        length_meters,
        cruise_speed,
        acceleration,
        deceleration,
    )
    reference_speed = fill_reference_series(
        numeric_column(frame, "ControlReferenceSpeedMetersPerSecond"),
        ideal_speed,
    )
    if np.any(has_control_reference):
        reference_speed[~has_control_reference] = np.nan
    motion_frame_yaw_radians = np.deg2rad(
        numeric_column(frame, "ReferenceMotionFrameYawDegrees", 0.0)
    )
    motion_direction_cosine = np.cos(motion_frame_yaw_radians)
    motion_direction_sine = np.sin(motion_frame_yaw_radians)
    state_body_vx = numeric_column(frame, "StateBodyVxMetersPerSecond")
    state_body_vy = numeric_column(frame, "StateBodyVyMetersPerSecond")
    velocity_valid = (
        numeric_column(frame, "StateVelocityEstimateValid", 0.0) > 0.5
    )
    state_body_vx[~velocity_valid] = np.nan
    state_body_vy[~velocity_valid] = np.nan
    state_motion_speed = (
        state_body_vx * motion_direction_cosine
        + state_body_vy * motion_direction_sine
    )
    has_velocity_diagnostics = (
        numeric_column(frame, "HasVelocityDiagnostics", 0.0) > 0.5
    )
    detour_speed = numeric_column(
        frame,
        "DetourEstimatedBodyVxMetersPerSecond",
    )
    detour_speed_valid = (
        has_velocity_diagnostics
        & (
            numeric_column(
                frame,
                "DetourVelocityEstimateValid",
                0.0,
            )
            > 0.5
        )
    )
    detour_speed[~detour_speed_valid] = np.nan
    wheel_raw_body_vx = numeric_column(
        frame,
        "WheelFeedbackRawBodyVxMetersPerSecond",
    )
    wheel_filtered_body_vx = numeric_column(
        frame,
        "WheelFeedbackFilteredBodyVxMetersPerSecond",
    )
    wheel_raw_body_vy = numeric_column(
        frame,
        "WheelFeedbackRawBodyVyMetersPerSecond",
    )
    wheel_filtered_body_vy = numeric_column(
        frame,
        "WheelFeedbackFilteredBodyVyMetersPerSecond",
    )
    wheel_speed_valid = (
        has_velocity_diagnostics
        & (
            numeric_column(
                frame,
                "WheelFeedbackVelocityEstimateValid",
                0.0,
            )
            > 0.5
        )
    )
    wheel_raw_speed = (
        wheel_raw_body_vx * motion_direction_cosine
        + wheel_raw_body_vy * motion_direction_sine
    )
    wheel_filtered_speed = (
        wheel_filtered_body_vx * motion_direction_cosine
        + wheel_filtered_body_vy * motion_direction_sine
    )

    # 兼容尚未记录轮速Vy的旧版β=0实验；非零β缺少Vy时不能伪造投影速度。
    body_x_motion = np.abs(motion_direction_sine) <= 1e-12
    missing_raw_projection = ~np.isfinite(wheel_raw_speed)
    missing_filtered_projection = ~np.isfinite(wheel_filtered_speed)
    wheel_raw_speed[body_x_motion & missing_raw_projection] = (
        wheel_raw_body_vx[body_x_motion & missing_raw_projection]
    )
    wheel_filtered_speed[
        body_x_motion & missing_filtered_projection
    ] = wheel_filtered_body_vx[
        body_x_motion & missing_filtered_projection
    ]
    wheel_raw_speed[~wheel_speed_valid] = np.nan
    wheel_filtered_speed[~wheel_speed_valid] = np.nan
    actual_speed = np.where(
        np.isfinite(wheel_filtered_speed),
        wheel_filtered_speed,
        state_motion_speed,
    )
    command_speed = numeric_column(frame, "CommandSpeed")

    has_steering_diagnostics = (
        numeric_column(frame, "HasSteeringDiagnostics", 0.0) > 0.5
    )
    steering_angles_degrees = {}
    for wheel_name in (
        "LeftFront",
        "LeftRear",
        "RightFront",
        "RightRear",
    ):
        values = numeric_column(
            frame,
            f"ActualSteer{wheel_name}Degrees",
        )
        values[~has_steering_diagnostics] = np.nan
        steering_angles_degrees[wheel_name] = values

    has_gcp_command = (
        numeric_column(frame, "HasGcpCommand", 0.0) > 0.5
    )
    front_gcp_degrees = np.rad2deg(
        numeric_column(frame, "CommandFrontGcpAngleRadians")
    )
    rear_gcp_degrees = np.rad2deg(
        numeric_column(frame, "CommandRearGcpAngleRadians")
    )
    front_gcp_degrees[~has_gcp_command] = np.nan
    rear_gcp_degrees[~has_gcp_command] = np.nan

    return {
        "frame": frame,
        "time": time_seconds,
        "actual_x": actual_x,
        "actual_y": actual_y,
        "valid_position": valid_position,
        "raw_x": raw_x_meters,
        "raw_y": raw_y_meters,
        "valid_raw_position": valid_raw_position,
        "detour_tick_raw": detour_tick_raw,
        "detour_l_step": detour_l_step,
        "detour_position_step": detour_position_step,
        "state_lateral_step": state_lateral_step,
        "raw_detour_jump": raw_detour_jump,
        "state_lateral_jump": state_lateral_jump,
        "jump_magnitude": jump_magnitude,
        "jump_indices": jump_indices,
        "reference_x": reference_x,
        "reference_y": reference_y,
        "valid_reference_position": valid_reference_position,
        "start": start,
        "end": end,
        "length": length_meters,
        "lateral_error": lateral_error,
        "heading_error": heading_error,
        "reference_speed": reference_speed,
        "actual_speed": actual_speed,
        "detour_speed": detour_speed,
        "wheel_raw_speed": wheel_raw_speed,
        "wheel_filtered_speed": wheel_filtered_speed,
        "command_speed": command_speed,
        "steering_angles_degrees": steering_angles_degrees,
        "front_gcp_degrees": front_gcp_degrees,
        "rear_gcp_degrees": rear_gcp_degrees,
        "controller_name": first_text(
            frame,
            "ControllerName",
            "NewController",
        ),
        "trajectory_name": first_text(
            frame,
            "TrajectoryName",
            "Trajectory",
        ),
    }


def finite_rmse(values: np.ndarray) -> float:
    """计算忽略无效样本后的均方根值。"""
    finite = values[np.isfinite(values)]
    return float(np.sqrt(np.mean(finite**2))) if finite.size else np.nan


def save_figure(
    fig: plt.Figure,
    destination: Path,
    show: bool,
) -> None:
    """保存并关闭一张实验图。"""
    fig.tight_layout(rect=(0.0, 0.0, 1.0, 0.97))
    fig.savefig(destination, dpi=300, bbox_inches="tight")
    if show:
        plt.show()
    plt.close(fig)


def plot_experiment(
    csv_path: Path,
    output_directory: Path,
    show: bool,
) -> list[Path]:
    """为单份新版控制器CSV生成一张包含六个子图的实验总图。"""
    data = load_experiment(csv_path)
    output_directory.mkdir(parents=True, exist_ok=True)
    title = f"{data['controller_name']} - {data['trajectory_name']}"
    fig, axes = plt.subplots(3, 2, figsize=(18.0, 16.0))
    fig.suptitle(title, fontsize=16)

    # 1. 期望轨迹与实际轨迹。
    axis = axes[0, 0]
    valid_position = data["valid_position"]
    valid_raw_position = data["valid_raw_position"]
    valid_reference_position = data["valid_reference_position"]
    jump_indices = data["jump_indices"]
    if np.count_nonzero(valid_reference_position) >= 2:
        axis.plot(
            data["reference_x"][valid_reference_position],
            data["reference_y"][valid_reference_position],
            "--",
            linewidth=2.0,
            label="控制器实际使用的参考轨迹",
        )
    else:
        axis.plot(
            [data["start"][0], data["end"][0]],
            [data["start"][1], data["end"][1]],
            "--",
            linewidth=2.0,
            label="参考起终点连线",
        )
    axis.plot(
        data["raw_x"][valid_raw_position],
        data["raw_y"][valid_raw_position],
        ":",
        color="tab:gray",
        linewidth=1.2,
        alpha=0.85,
        label="Detour原始轨迹",
    )
    axis.plot(
        data["actual_x"][valid_position],
        data["actual_y"][valid_position],
        color="tab:orange",
        linewidth=1.5,
        label="控制使用的状态轨迹",
    )
    if jump_indices.size:
        axis.scatter(
            data["raw_x"][jump_indices],
            data["raw_y"][jump_indices],
            color="red",
            marker="x",
            s=65,
            linewidths=1.8,
            zorder=8,
            label="疑似定位/状态突变",
        )
    axis.scatter(*data["start"], color="green", s=45, label="起点")
    axis.scatter(*data["end"], color="red", s=45, label="终点")
    # 诊断图优先展示厘米级横向变化；横纵轴独立缩放，避免4m行程
    # 将数厘米的定位阶跃压缩成几乎不可见的一条细线。
    axis.set_aspect("auto")
    axis.set_xlabel("世界坐标X / m")
    axis.set_ylabel("世界坐标Y / m")
    axis.set_title("期望轨迹与状态轨迹对比（横纵轴独立缩放）")
    axis.grid(True, alpha=0.3)
    axis.legend(fontsize=7, loc="upper left")

    if jump_indices.size:
        strongest_jump_index = int(
            jump_indices[
                np.argmax(
                    np.abs(
                        data["jump_magnitude"][jump_indices]
                    )
                )
            ]
        )
        context_start = max(
            0,
            strongest_jump_index - JUMP_INSET_CONTEXT_SAMPLES,
        )
        context_end = min(
            len(data["time"]),
            strongest_jump_index + JUMP_INSET_CONTEXT_SAMPLES + 1,
        )
        context = np.arange(context_start, context_end)
        inset = axis.inset_axes([0.54, 0.08, 0.43, 0.43])
        inset.set_zorder(10)
        inset.set_facecolor("white")
        context_reference_valid = (
            data["valid_reference_position"][context]
        )
        if np.count_nonzero(context_reference_valid) >= 2:
            reference_context = context[context_reference_valid]
            inset.plot(
                data["reference_x"][reference_context],
                data["reference_y"][reference_context],
                "--",
                linewidth=1.2,
                color="tab:blue",
            )
        context_raw_valid = data["valid_raw_position"][context]
        raw_context = context[context_raw_valid]
        inset.plot(
            data["raw_x"][raw_context],
            data["raw_y"][raw_context],
            ":",
            linewidth=1.0,
            color="tab:gray",
        )
        context_state_valid = data["valid_position"][context]
        state_context = context[context_state_valid]
        inset.plot(
            data["actual_x"][state_context],
            data["actual_y"][state_context],
            linewidth=1.2,
            color="tab:orange",
        )
        inset.scatter(
            data["raw_x"][strongest_jump_index],
            data["raw_y"][strongest_jump_index],
            color="red",
            marker="x",
            s=45,
            linewidths=1.5,
            zorder=8,
        )
        jump_descriptions = []
        if data["raw_detour_jump"][strongest_jump_index]:
            jump_descriptions.append(
                "Detour位移="
                f"{data['detour_position_step'][strongest_jump_index] * 1000.0:.1f}mm"
            )
        if data["state_lateral_jump"][strongest_jump_index]:
            jump_descriptions.append(
                "状态横向Δ="
                f"{data['state_lateral_step'][strongest_jump_index] * 1000.0:+.1f}mm"
            )
        diagnostic_parts = []
        detour_tick = data["detour_tick_raw"][strongest_jump_index]
        if detour_tick:
            diagnostic_parts.append(f"tick={detour_tick}")
        detour_l_step = data["detour_l_step"][strongest_jump_index]
        if np.isfinite(detour_l_step):
            diagnostic_parts.append(f"l_step={detour_l_step:g}")
        diagnostic_suffix = (
            "\n" + "  ".join(diagnostic_parts)
            if diagnostic_parts
            else ""
        )
        inset.set_title(
            f"最大疑似突变：t={data['time'][strongest_jump_index]:.3f}s\n"
            f"{'，'.join(jump_descriptions)}"
            f"{diagnostic_suffix}",
            fontsize=7,
        )
        inset.set_aspect("auto")
        inset.tick_params(labelsize=6)
        inset.grid(True, alpha=0.25)

    # 2. 横向误差。
    lateral_mm = data["lateral_error"] * 1000.0
    lateral_rmse_mm = finite_rmse(lateral_mm)
    axis = axes[0, 1]
    axis.plot(data["time"], lateral_mm, linewidth=1.5)
    if jump_indices.size:
        for jump_index in jump_indices:
            axis.axvline(
                data["time"][jump_index],
                color="red",
                linewidth=0.8,
                alpha=0.35,
            )
        valid_jump_error = (
            data["state_lateral_jump"][jump_indices]
            & np.isfinite(lateral_mm[jump_indices])
        )
        visible_jump_indices = jump_indices[valid_jump_error]
        if visible_jump_indices.size:
            axis.scatter(
                data["time"][visible_jump_indices],
                lateral_mm[visible_jump_indices],
                color="red",
                marker="x",
                s=45,
                linewidths=1.5,
                zorder=7,
                label="控制状态横向突变",
            )
    axis.axhline(0.0, color="black", linewidth=0.8)
    axis.set_xlabel("时间 / s")
    axis.set_ylabel("横向误差 / mm")
    axis.set_title(
        "横向误差（轨迹在车辆左侧为正）\n"
        f"RMSE={lateral_rmse_mm:.2f}mm"
    )
    axis.grid(True, alpha=0.3)
    if jump_indices.size and np.any(
        data["state_lateral_jump"][jump_indices]
        & np.isfinite(lateral_mm[jump_indices])
    ):
        axis.legend(fontsize=8)

    # 3. 航向误差。
    heading_degrees = np.rad2deg(data["heading_error"])
    heading_rmse_degrees = finite_rmse(heading_degrees)
    axis = axes[1, 0]
    axis.plot(data["time"], heading_degrees, linewidth=1.5)
    axis.axhline(0.0, color="black", linewidth=0.8)
    axis.set_xlabel("时间 / s")
    axis.set_ylabel("航向角偏差 / °")
    axis.set_title(
        "航向角偏差：参考轨迹航向－实际车体航向（逆时针为正）\n"
        f"RMSE={heading_rmse_degrees:.3f}°"
    )
    axis.grid(True, alpha=0.3)

    # 4. 参考、命令、Detour车头分量和沿β投影的轮速解算速度。
    speed_error = data["actual_speed"] - data["reference_speed"]
    speed_rmse = finite_rmse(speed_error)
    axis = axes[1, 1]
    axis.plot(
        data["time"],
        data["reference_speed"],
        linewidth=1.8,
        label="控制器实际参考速度",
    )
    axis.plot(
        data["time"],
        data["command_speed"],
        "--",
        linewidth=1.3,
        label="纵向控制器下发速度",
    )
    if np.any(np.isfinite(data["detour_speed"])):
        axis.plot(
            data["time"],
            data["detour_speed"],
            ":",
            linewidth=1.2,
            label="Detour估计Vx（车头分量）",
        )
    if np.any(np.isfinite(data["wheel_filtered_speed"])):
        wheel_filtered_valid = np.isfinite(
            data["wheel_filtered_speed"]
        )
        axis.scatter(
            data["time"][wheel_filtered_valid],
            data["wheel_filtered_speed"][wheel_filtered_valid],
            color="tab:red",
            s=14,
            marker="o",
            zorder=5,
            label="轮速解算β方向速度（控制使用）",
        )
    else:
        axis.plot(
            data["time"],
            data["actual_speed"],
            linewidth=1.5,
            label="控制器实际纵向速度",
        )
    axis.set_xlabel("时间 / s")
    axis.set_ylabel("速度 / (m/s)")
    axis.set_title(
        "参考速度、控制命令与观测速度\n"
        f"轮速β方向速度相对参考速度RMSE={speed_rmse:.4f}m/s"
    )
    axis.grid(True, alpha=0.3)
    axis.legend(fontsize=8)

    # 5. 四个舵轮的实际机械转角。
    axis = axes[2, 0]
    wheel_labels = {
        "LeftFront": "左前轮",
        "LeftRear": "左后轮",
        "RightFront": "右前轮",
        "RightRear": "右后轮",
    }
    steering_data_available = False
    for wheel_name, wheel_label in wheel_labels.items():
        wheel_angles = data["steering_angles_degrees"][wheel_name]
        if np.any(np.isfinite(wheel_angles)):
            steering_data_available = True
            axis.plot(
                data["time"],
                wheel_angles,
                linewidth=1.2,
                label=wheel_label,
            )
    if steering_data_available:
        axis.axhline(0.0, color="black", linewidth=0.8)
        axis.legend(fontsize=8, ncol=2)
    else:
        axis.text(
            0.5,
            0.5,
            "CSV不含四舵轮转角诊断数据",
            ha="center",
            va="center",
            transform=axis.transAxes,
        )
    axis.set_xlabel("时间 / s")
    axis.set_ylabel("实际舵角 / °")
    axis.set_title("四个舵轮实际反馈转角")
    axis.grid(True, alpha=0.3)

    # 6. 经过角速度限制后实际发送的前、后虚拟GCP转角。
    axis = axes[2, 1]
    gcp_data_available = (
        np.any(np.isfinite(data["front_gcp_degrees"]))
        or np.any(np.isfinite(data["rear_gcp_degrees"]))
    )
    if gcp_data_available:
        axis.plot(
            data["time"],
            data["front_gcp_degrees"],
            linewidth=1.4,
            label="前GCP",
        )
        axis.plot(
            data["time"],
            data["rear_gcp_degrees"],
            linewidth=1.4,
            label="后GCP",
        )
        axis.axhline(0.0, color="black", linewidth=0.8)
        axis.legend(fontsize=8)
    else:
        axis.text(
            0.5,
            0.5,
            "CSV不含前后GCP转角数据",
            ha="center",
            va="center",
            transform=axis.transAxes,
        )
    axis.set_xlabel("时间 / s")
    axis.set_ylabel("GCP命令角 / °")
    axis.set_title("前后虚拟GCP实际发送转角")
    axis.grid(True, alpha=0.3)

    destination = output_directory / f"{csv_path.stem}_summary_6plots.png"
    save_figure(fig, destination, show)

    print(
        f"{csv_path.name}: 横向RMSE={lateral_rmse_mm:.3f}mm, "
        f"航向RMSE={heading_rmse_degrees:.4f}°, "
        f"速度RMSE={speed_rmse:.5f}m/s"
    )
    if jump_indices.size:
        strongest_jump_index = int(
            jump_indices[
                np.argmax(
                    np.abs(
                        data["jump_magnitude"][jump_indices]
                    )
                )
            ]
        )
        print(
            f"  检出{jump_indices.size}个疑似定位/状态突变，"
            f"最大幅值={data['jump_magnitude'][strongest_jump_index] * 1000.0:.2f}mm，"
            f"时刻={data['time'][strongest_jump_index]:.3f}s"
        )
    print(f"已生成六子图总图：{destination}")
    return [destination]


def discover_csv_files(arguments: list[str]) -> list[Path]:
    """读取命令行文件或目录；目录中只选取非计时CSV。"""
    if arguments:
        files = []
        for item in arguments:
            path = Path(item).expanduser().resolve()
            if path.is_dir():
                files.extend(
                    sorted(
                        candidate
                        for candidate in path.glob("*.csv")
                        if not candidate.stem.endswith("_timing")
                    )
                )
            else:
                files.append(path)
    else:
        files = sorted(
            path
            for path in SCRIPT_DIR.glob("*.csv")
            if not path.stem.endswith("_timing")
        )
        files.extend(
            sorted(
                path
                for path in (SCRIPT_DIR / "data").glob("*.csv")
                if not path.stem.endswith("_timing")
            )
        )
    files = list(dict.fromkeys(path for path in files if path.is_file()))
    if not files:
        raise FileNotFoundError(
            "没有找到轨迹CSV；请传入文件、目录，或将文件放到脚本目录/data中。"
        )
    return files


def main() -> None:
    """解析命令行并批量处理新版控制器实验CSV。"""
    parser = argparse.ArgumentParser(
        description="绘制新版控制器轨迹实验的六子图总图。"
    )
    parser.add_argument(
        "csv",
        nargs="*",
        help="需要处理的轨迹CSV文件或包含轨迹CSV的目录。",
    )
    parser.add_argument(
        "--output-dir",
        help="图片输出目录；默认使用脚本目录/plots。",
    )
    parser.add_argument(
        "--show",
        action="store_true",
        help="保存图片后同时显示窗口。",
    )
    arguments = parser.parse_args()
    configure_matplotlib()

    output_directory = (
        Path(arguments.output_dir).expanduser().resolve()
        if arguments.output_dir
        else SCRIPT_DIR / "plots"
    )
    failed = 0
    for csv_path in discover_csv_files(arguments.csv):
        try:
            plot_experiment(csv_path, output_directory, arguments.show)
        except Exception as exception:
            failed += 1
            print(f"处理失败：{csv_path}：{exception}")

    if failed:
        raise SystemExit(f"共有{failed}个CSV处理失败。")


if __name__ == "__main__":
    main()