app.py

import os
import json
import pandas as pd
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
import gradio as gr

# 加载数据
def load_json(filename):
    possible_paths = [
        f'results/{filename}',
        filename,
        os.path.join(os.path.dirname(__file__), 'results', filename),
        os.path.join(os.path.dirname(__file__), filename)
    ]
    for path in possible_paths:
        if os.path.exists(path):
            with open(path, 'r', encoding='utf-8') as f:
                return json.load(f)
    raise FileNotFoundError(f"Cannot find {filename}")

# 加载所有数据
va_data = load_json('va_models.json')
vla_data = load_json('vla_models.json')
datasets_data = load_json('datasets.json')
nuscenes_data = load_json('nuscenes_results.json')
wod_data = load_json('wod_results.json')
navsim_data = load_json('navsim_results.json')
bench2drive_data = load_json('bench2drive_results.json')

df_va = pd.DataFrame(va_data)
df_vla = pd.DataFrame(vla_data)
df_datasets = pd.DataFrame(datasets_data)
df_nuscenes = pd.DataFrame(nuscenes_data)
df_wod = pd.DataFrame(wod_data)
df_navsim = pd.DataFrame(navsim_data)
df_bench2drive = pd.DataFrame(bench2drive_data)

INPUT_ICONS = {'camera': '📷', 'lidar': '🔍', 'status': '⚙️', 'prompt': '💭', 
               'instruction': '📝', 'scene': '🌆', 'traffic': '🚦', 'context': '📚'}

ACTION_LABELS = {'RL': 'Policy w/ RL', 'REG': 'Decoder+MLP', 'SEL': 'Traj. Selection',
                'GEN': 'Generative Model', 'LH': 'Language Head'}

# 格式化函数
def format_inputs(inputs):
    if isinstance(inputs, list):
        return ' '.join([INPUT_ICONS.get(inp, inp) for inp in inputs])
    return str(inputs)

def format_datasets(datasets):
    if isinstance(datasets, list):
        return ', '.join(datasets[:3])
    return str(datasets)

# VA Models
def update_va_table(category, venue, action, search):
    df = df_va.copy()
    if category != "All":
        df = df[df['category'] == category]
    if venue:
        df = df[df['venue'].str.contains(venue, case=False, na=False)]
    if action != "All":
        df = df[df['action'] == action]
    if search:
        mask = df['model'].str.contains(search, case=False, na=False) | df['vision'].str.contains(search, case=False, na=False)
        df = df[mask]
    display_df = df[['id', 'model', 'venue', 'input', 'dataset', 'vision', 'action', 'output', 'category']].copy()
    display_df['input'] = display_df['input'].apply(format_inputs)
    display_df['dataset'] = display_df['dataset'].apply(format_datasets)
    display_df['action'] = display_df['action'].map(ACTION_LABELS)
    return display_df, f"**📊 Statistics:** Total {len(df)} models | Categories: {df['category'].nunique()}"

# 初始化VA表格数据
def init_va_table():
    return update_va_table("All", "", "All", "")

def create_va_plot(category):
    df = df_va.copy()
    if category != "All":
        df = df[df['category'] == category]
    bg_color, panel_color = "#0e1117", "#161b22"
    bar_colors = ['#4cc9f0', '#f72585', '#7209b7', '#3a0ca3', '#4361ee', '#4895ef', '#560bad']
    text_color, grid_color = "#c9d1d9", "#30363d"
    
    fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(14, 5))
    fig.patch.set_facecolor(bg_color)
    
    category_counts = df['category'].value_counts()
    ax1.barh(category_counts.index, category_counts.values, color=bar_colors[:len(category_counts)])
    ax1.set_facecolor(panel_color)
    ax1.set_xlabel('Count', color=text_color, fontsize=11)
    ax1.set_title('Models by Category', color=text_color, fontsize=13, fontweight='bold')
    ax1.tick_params(colors=text_color)
    ax1.grid(axis='x', linestyle='--', alpha=0.3, color=grid_color)
    ax1.xaxis.set_major_locator(plt.MaxNLocator(integer=True))
    for spine in ['top', 'right', 'left']:
        ax1.spines[spine].set_visible(False)
    ax1.spines['bottom'].set_color(grid_color)
    
    action_counts = df['action'].value_counts()
    ax2.bar(range(len(action_counts)), action_counts.values, color=bar_colors[:len(action_counts)])
    ax2.set_facecolor(panel_color)
    ax2.set_xticks(range(len(action_counts)))
    ax2.set_xticklabels([ACTION_LABELS.get(a, a) for a in action_counts.index], rotation=45, ha='right')
    ax2.set_ylabel('Count', color=text_color, fontsize=11)
    ax2.set_title('Models by Action Type', color=text_color, fontsize=13, fontweight='bold')
    ax2.tick_params(colors=text_color)
    ax2.grid(axis='y', linestyle='--', alpha=0.3, color=grid_color)
    ax2.yaxis.set_major_locator(plt.MaxNLocator(integer=True))
    for spine in ['top', 'right']:
        ax2.spines[spine].set_visible(False)
    ax2.spines['left'].set_color(grid_color)
    ax2.spines['bottom'].set_color(grid_color)
    
    plt.tight_layout()
    result = fig
    plt.close('all')
    return result

# VLA Models
def update_vla_table(category, venue, language, search):
    df = df_vla.copy()
    if category != "All":
        df = df[df['category'] == category]
    if venue:
        df = df[df['venue'].str.contains(venue, case=False, na=False)]
    if language:
        df = df[df['language'].str.contains(language, case=False, na=False)]
    if search:
        mask = df['model'].str.contains(search, case=False, na=False) | df['vision'].str.contains(search, case=False, na=False) | df['language'].str.contains(search, case=False, na=False)
        df = df[mask]
    display_df = df[['id', 'model', 'venue', 'input', 'dataset', 'vision', 'language', 'action', 'output', 'category']].copy()
    display_df['input'] = display_df['input'].apply(format_inputs)
    display_df['dataset'] = display_df['dataset'].apply(format_datasets)
    display_df['action'] = display_df['action'].map(ACTION_LABELS)
    return display_df, f"**📊 Statistics:** Total {len(df)} models | Categories: {df['category'].nunique()}"

# 初始化VLA表格数据
def init_vla_table():
    return update_vla_table("All", "", "", "")

def create_vla_plot(category):
    df = df_vla.copy()
    if category != "All":
        df = df[df['category'] == category]
    bg_color, panel_color = "#0e1117", "#161b22"
    bar_colors = ['#4cc9f0', '#f72585', '#7209b7', '#3a0ca3']
    text_color, grid_color = "#c9d1d9", "#30363d"
    
    fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(14, 5))
    fig.patch.set_facecolor(bg_color)
    
    category_counts = df['category'].value_counts()
    ax1.barh(category_counts.index, category_counts.values, color=bar_colors[:len(category_counts)])
    ax1.set_facecolor(panel_color)
    ax1.set_xlabel('Count', color=text_color, fontsize=11)
    ax1.set_title('VLA Models by Category', color=text_color, fontsize=13, fontweight='bold')
    ax1.tick_params(colors=text_color)
    ax1.grid(axis='x', linestyle='--', alpha=0.3, color=grid_color)
    ax1.xaxis.set_major_locator(plt.MaxNLocator(integer=True))
    for spine in ['top', 'right', 'left']:
        ax1.spines[spine].set_visible(False)
    ax1.spines['bottom'].set_color(grid_color)
    
    language_counts = df['language'].value_counts().head(10)
    ax2.bar(range(len(language_counts)), language_counts.values, color=bar_colors[:len(language_counts)])
    ax2.set_facecolor(panel_color)
    ax2.set_xticks(range(len(language_counts)))
    ax2.set_xticklabels(language_counts.index, rotation=45, ha='right')
    ax2.set_ylabel('Count', color=text_color, fontsize=11)
    ax2.set_title('Top Language Models', color=text_color, fontsize=13, fontweight='bold')
    ax2.tick_params(colors=text_color)
    ax2.grid(axis='y', linestyle='--', alpha=0.3, color=grid_color)
    ax2.yaxis.set_major_locator(plt.MaxNLocator(integer=True))
    for spine in ['top', 'right']:
        ax2.spines[spine].set_visible(False)
    ax2.spines['left'].set_color(grid_color)
    ax2.spines['bottom'].set_color(grid_color)
    
    plt.tight_layout()
    result = fig
    plt.close('all')
    return result

# Datasets
def update_datasets_table(category, year, search):
    df = df_datasets.copy()
    if category != "All":
        df = df[df['category'] == category]
    if year:
        try:
            df = df[df['year'] == int(year)]
        except:
            pass
    if search:
        mask = df['dataset'].str.contains(search, case=False, na=False) | df['sensor'].str.contains(search, case=False, na=False)
        df = df[mask]
    return df, f"**📊 Statistics:** Total {len(df)} datasets | Years: {df['year'].nunique()}"

# 初始化Datasets表格数据
def init_datasets_table():
    return update_datasets_table("All", "", "")

# Benchmark tables update functions
def update_nuscenes_table(category, search):
    df = df_nuscenes.copy()
    if category != "All":
        df = df[df['category'] == category]
    if search:
        mask = df['model'].str.contains(search, case=False, na=False)
        df = df[mask]
    return df, f"**📊 Statistics:** Total {len(df)} models"

# 初始化nuScenes表格数据
def init_nuscenes_table():
    return update_nuscenes_table("All", "")

def create_nuscenes_plot(category):
    """创建nuScenes统计图表"""
    df = df_nuscenes.copy()
    if category != "All":
        df = df[df['category'] == category]
    
    bg_color, panel_color = "#0e1117", "#161b22"
    bar_colors = ['#4cc9f0', '#f72585', '#7209b7', '#3a0ca3']
    text_color, grid_color = "#c9d1d9", "#30363d"
    
    fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(14, 5))
    fig.patch.set_facecolor(bg_color)
    
    # 1. 按类别统计
    category_counts = df['category'].value_counts()
    ax1.bar(range(len(category_counts)), category_counts.values, color=bar_colors[:len(category_counts)])
    ax1.set_facecolor(panel_color)
    ax1.set_xticks(range(len(category_counts)))
    ax1.set_xticklabels(category_counts.index, rotation=15, ha='right')
    ax1.set_ylabel('Count', color=text_color, fontsize=11)
    ax1.set_title('Models by Category', color=text_color, fontsize=13, fontweight='bold')
    ax1.tick_params(colors=text_color)
    ax1.grid(axis='y', linestyle='--', alpha=0.3, color=grid_color)
    ax1.yaxis.set_major_locator(plt.MaxNLocator(integer=True))
    for spine in ['top', 'right']:
        ax1.spines[spine].set_visible(False)
    ax1.spines['left'].set_color(grid_color)
    ax1.spines['bottom'].set_color(grid_color)
    
    # 2. Top 10 模型按L2 Avg排序
    top_models = df.nsmallest(10, 'l2_avg')[['model', 'l2_avg']].dropna()
    ax2.barh(range(len(top_models)), top_models['l2_avg'].values, color=bar_colors[2])
    ax2.set_facecolor(panel_color)
    ax2.set_yticks(range(len(top_models)))
    ax2.set_yticklabels(top_models['model'].values, fontsize=9)
    ax2.set_xlabel('L2 Avg (m)', color=text_color, fontsize=11)
    ax2.set_title('Top 10 Models by L2 Error', color=text_color, fontsize=13, fontweight='bold')
    ax2.tick_params(colors=text_color)
    ax2.grid(axis='x', linestyle='--', alpha=0.3, color=grid_color)
    ax2.invert_yaxis()
    for spine in ['top', 'right', 'left']:
        ax2.spines[spine].set_visible(False)
    ax2.spines['bottom'].set_color(grid_color)
    
    plt.tight_layout()
    result = fig
    plt.close('all')
    return result

def update_wod_table(category, search):
    df = df_wod.copy()
    if category != "All":
        df = df[df['category'] == category]
    if search:
        mask = df['model'].str.contains(search, case=False, na=False)
        df = df[mask]
    return df, f"**📊 Statistics:** Total {len(df)} models"

# 初始化WOD表格数据
def init_wod_table():
    return update_wod_table("All", "")

def create_wod_plot(category):
    """创建WOD统计图表"""
    df = df_wod.copy()
    if category != "All":
        df = df[df['category'] == category]
    
    bg_color, panel_color = "#0e1117", "#161b22"
    bar_colors = ['#4cc9f0', '#f72585', '#7209b7', '#3a0ca3']
    text_color, grid_color = "#c9d1d9", "#30363d"
    
    fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(14, 5))
    fig.patch.set_facecolor(bg_color)
    
    # 1. 按类别统计
    category_counts = df['category'].value_counts()
    ax1.bar(range(len(category_counts)), category_counts.values, color=bar_colors[:len(category_counts)])
    ax1.set_facecolor(panel_color)
    ax1.set_xticks(range(len(category_counts)))
    ax1.set_xticklabels(category_counts.index, rotation=15, ha='right')
    ax1.set_ylabel('Count', color=text_color, fontsize=11)
    ax1.set_title('Models by Category', color=text_color, fontsize=13, fontweight='bold')
    ax1.tick_params(colors=text_color)
    ax1.grid(axis='y', linestyle='--', alpha=0.3, color=grid_color)
    ax1.yaxis.set_major_locator(plt.MaxNLocator(integer=True))
    for spine in ['top', 'right']:
        ax1.spines[spine].set_visible(False)
    ax1.spines['left'].set_color(grid_color)
    ax1.spines['bottom'].set_color(grid_color)
    
    # 2. Top模型按RFS Overall排序
    top_models = df.nlargest(10, 'rfs_overall')[['model', 'rfs_overall']].dropna()
    if len(top_models) > 0:
        ax2.barh(range(len(top_models)), top_models['rfs_overall'].values, color=bar_colors[1])
        ax2.set_facecolor(panel_color)
        ax2.set_yticks(range(len(top_models)))
        ax2.set_yticklabels(top_models['model'].values, fontsize=9)
        ax2.set_xlabel('RFS Overall', color=text_color, fontsize=11)
        ax2.set_title('Top Models by RFS Overall', color=text_color, fontsize=13, fontweight='bold')
        ax2.tick_params(colors=text_color)
        ax2.grid(axis='x', linestyle='--', alpha=0.3, color=grid_color)
        ax2.invert_yaxis()
    else:
        ax2.text(0.5, 0.5, 'No data available', ha='center', va='center', 
                color=text_color, fontsize=12, transform=ax2.transAxes)
        ax2.set_facecolor(panel_color)
    
    for spine in ['top', 'right', 'left']:
        ax2.spines[spine].set_visible(False)
    ax2.spines['bottom'].set_color(grid_color)
    
    plt.tight_layout()
    result = fig
    plt.close('all')
    return result

def update_navsim_table(category, search):
    df = df_navsim.copy()
    if category != "All":
        df = df[df['category'] == category]
    if search:
        mask = df['model'].str.contains(search, case=False, na=False)
        df = df[mask]
    return df, f"**📊 Statistics:** Total {len(df)} models"

# 初始化NAVSIM表格数据
def init_navsim_table():
    return update_navsim_table("All", "")

def create_navsim_plot(category):
    """创建NAVSIM统计图表"""
    df = df_navsim.copy()
    if category != "All":
        df = df[df['category'] == category]
    
    bg_color, panel_color = "#0e1117", "#161b22"
    bar_colors = ['#4cc9f0', '#f72585', '#7209b7', '#3a0ca3']
    text_color, grid_color = "#c9d1d9", "#30363d"
    
    fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(14, 5))
    fig.patch.set_facecolor(bg_color)
    
    # 1. 按类别统计
    category_counts = df['category'].value_counts()
    ax1.bar(range(len(category_counts)), category_counts.values, color=bar_colors[:len(category_counts)])
    ax1.set_facecolor(panel_color)
    ax1.set_xticks(range(len(category_counts)))
    ax1.set_xticklabels(category_counts.index, rotation=15, ha='right')
    ax1.set_ylabel('Count', color=text_color, fontsize=11)
    ax1.set_title('Models by Category', color=text_color, fontsize=13, fontweight='bold')
    ax1.tick_params(colors=text_color)
    ax1.grid(axis='y', linestyle='--', alpha=0.3, color=grid_color)
    ax1.yaxis.set_major_locator(plt.MaxNLocator(integer=True))
    for spine in ['top', 'right']:
        ax1.spines[spine].set_visible(False)
    ax1.spines['left'].set_color(grid_color)
    ax1.spines['bottom'].set_color(grid_color)
    
    # 2. Top 10 模型按PDMS排序
    top_models = df.nlargest(10, 'pdms')[['model', 'pdms']]
    ax2.barh(range(len(top_models)), top_models['pdms'].values, color=bar_colors[0])
    ax2.set_facecolor(panel_color)
    ax2.set_yticks(range(len(top_models)))
    ax2.set_yticklabels(top_models['model'].values, fontsize=9)
    ax2.set_xlabel('PDMS Score', color=text_color, fontsize=11)
    ax2.set_title('Top 10 Models by PDMS', color=text_color, fontsize=13, fontweight='bold')
    ax2.tick_params(colors=text_color)
    ax2.grid(axis='x', linestyle='--', alpha=0.3, color=grid_color)
    ax2.invert_yaxis()
    for spine in ['top', 'right', 'left']:
        ax2.spines[spine].set_visible(False)
    ax2.spines['bottom'].set_color(grid_color)
    
    plt.tight_layout()
    result = fig
    plt.close('all')
    return result

def update_bench2drive_table(category, search):
    df = df_bench2drive.copy()
    if category != "All":
        df = df[df['category'] == category]
    if search:
        mask = df['model'].str.contains(search, case=False, na=False)
        df = df[mask]
    return df, f"**📊 Statistics:** Total {len(df)} models"

# 初始化Bench2Drive表格数据
def init_bench2drive_table():
    return update_bench2drive_table("All", "")

def create_bench2drive_plot(category):
    """创建Bench2Drive统计图表"""
    df = df_bench2drive.copy()
    if category != "All":
        df = df[df['category'] == category]
    
    bg_color, panel_color = "#0e1117", "#161b22"
    bar_colors = ['#4cc9f0', '#f72585', '#7209b7', '#3a0ca3']
    text_color, grid_color = "#c9d1d9", "#30363d"
    
    fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(14, 5))
    fig.patch.set_facecolor(bg_color)
    
    # 1. 按类别统计
    category_counts = df['category'].value_counts()
    ax1.bar(range(len(category_counts)), category_counts.values, color=bar_colors[:len(category_counts)])
    ax1.set_facecolor(panel_color)
    ax1.set_xticks(range(len(category_counts)))
    ax1.set_xticklabels(category_counts.index, rotation=15, ha='right')
    ax1.set_ylabel('Count', color=text_color, fontsize=11)
    ax1.set_title('Models by Category', color=text_color, fontsize=13, fontweight='bold')
    ax1.tick_params(colors=text_color)
    ax1.grid(axis='y', linestyle='--', alpha=0.3, color=grid_color)
    ax1.yaxis.set_major_locator(plt.MaxNLocator(integer=True))
    for spine in ['top', 'right']:
        ax1.spines[spine].set_visible(False)
    ax1.spines['left'].set_color(grid_color)
    ax1.spines['bottom'].set_color(grid_color)
    
    # 2. Top 10 模型按DS排序
    top_models = df.nlargest(10, 'ds')[['model', 'ds']]
    ax2.barh(range(len(top_models)), top_models['ds'].values, color=bar_colors[3])
    ax2.set_facecolor(panel_color)
    ax2.set_yticks(range(len(top_models)))
    ax2.set_yticklabels(top_models['model'].values, fontsize=9)
    ax2.set_xlabel('Driving Score', color=text_color, fontsize=11)
    ax2.set_title('Top 10 Models by DS', color=text_color, fontsize=13, fontweight='bold')
    ax2.tick_params(colors=text_color)
    ax2.grid(axis='x', linestyle='--', alpha=0.3, color=grid_color)
    ax2.invert_yaxis()
    for spine in ['top', 'right', 'left']:
        ax2.spines[spine].set_visible(False)
    ax2.spines['bottom'].set_color(grid_color)
    
    plt.tight_layout()
    result = fig
    plt.close('all')
    return result

# 预生成初始数据和图表
initial_va_df, initial_va_stats = init_va_table()
initial_vla_df, initial_vla_stats = init_vla_table()
initial_datasets_df, initial_datasets_stats = init_datasets_table()
initial_nuscenes_df, initial_nuscenes_stats = init_nuscenes_table()
initial_wod_df, initial_wod_stats = init_wod_table()
initial_navsim_df, initial_navsim_stats = init_navsim_table()
initial_bench2drive_df, initial_bench2drive_stats = init_bench2drive_table()

initial_va_plot = create_va_plot("All")
initial_vla_plot = create_vla_plot("All")
initial_nuscenes_plot = create_nuscenes_plot("All")
initial_wod_plot = create_wod_plot("All")
initial_navsim_plot = create_navsim_plot("All")
initial_bench2drive_plot = create_bench2drive_plot("All")

# Gradio UI
with gr.Blocks(css="#title {text-align: center;} .gradio-container {max-width: 100% !important;}") as demo:
    gr.Markdown("# 🚗 VLA for Autonomous Driving: Model Leaderboard", elem_id="title")
    gr.Markdown("### 📄 Survey: *Vision-Language-Action Models for Autonomous Driving*")
    
    with gr.Tabs():
        # Tab 1: VA Models
        with gr.Tab("Vision-Action Models (VA)"):
            gr.Markdown("### Table 1: Vision-Action Models in Autonomous Driving")
            va_stats = gr.Markdown(initial_va_stats)
            with gr.Row():
                va_category = gr.Dropdown(label="Category", choices=["All"] + sorted(df_va['category'].unique().tolist()), value="All")
                va_venue = gr.Textbox(label="Filter by Venue", placeholder="e.g., CVPR, ICCV...")
                va_action = gr.Dropdown(label="Action Type", choices=["All"] + sorted(df_va['action'].unique().tolist()), value="All")
                va_search = gr.Textbox(label="Search", placeholder="e.g., TransFuser...")
            va_update_btn = gr.Button("🔍 Update", variant="primary")
            va_table = gr.Dataframe(value=initial_va_df, label="VA Models", interactive=False, wrap=True)
            va_plot = gr.Plot(label="Statistics", value=initial_va_plot, format="png")
            va_update_btn.click(fn=lambda cat, ven, act, search: (*update_va_table(cat, ven, act, search), create_va_plot(cat)),
                               inputs=[va_category, va_venue, va_action, va_search], outputs=[va_table, va_stats, va_plot])
        
        # Tab 2: VLA Models
        with gr.Tab("Vision-Language-Action Models (VLA)"):
            gr.Markdown("### Table 3: Vision-Language-Action Models in Autonomous Driving")
            vla_stats = gr.Markdown(initial_vla_stats)
            with gr.Row():
                vla_category = gr.Dropdown(label="Category", choices=["All"] + sorted(df_vla['category'].unique().tolist()), value="All")
                vla_venue = gr.Textbox(label="Filter by Venue", placeholder="e.g., CVPR...")
                vla_language = gr.Textbox(label="Filter by Language Model", placeholder="e.g., Qwen...")
                vla_search = gr.Textbox(label="Search", placeholder="e.g., AutoVLA...")
            vla_update_btn = gr.Button("🔍 Update", variant="primary")
            vla_table = gr.Dataframe(value=initial_vla_df, label="VLA Models", interactive=False, wrap=True)
            vla_plot = gr.Plot(label="Statistics", value=initial_vla_plot, format="png")
            vla_update_btn.click(fn=lambda cat, ven, lang, search: (*update_vla_table(cat, ven, lang, search), create_vla_plot(cat)),
                                inputs=[vla_category, vla_venue, vla_language, vla_search], outputs=[vla_table, vla_stats, vla_plot])
        
        # Tab 3: Datasets
        with gr.Tab("Datasets & Benchmarks"):
            gr.Markdown("### Table 4: Summary of Datasets & Benchmarks")
            datasets_stats = gr.Markdown(initial_datasets_stats)
            with gr.Row():
                datasets_category = gr.Dropdown(label="Category", choices=["All"] + sorted(df_datasets['category'].unique().tolist()), value="All")
                datasets_year = gr.Textbox(label="Filter by Year", placeholder="e.g., 2024...")
                datasets_search = gr.Textbox(label="Search", placeholder="e.g., nuScenes...")
            datasets_update_btn = gr.Button("🔍 Update", variant="primary")
            datasets_table = gr.Dataframe(value=initial_datasets_df, label="Datasets", interactive=False, wrap=True)
            datasets_update_btn.click(fn=update_datasets_table, inputs=[datasets_category, datasets_year, datasets_search], outputs=[datasets_table, datasets_stats])
        
        # Tab 4: nuScenes
        with gr.Tab("nuScenes Open-Loop"):
            gr.Markdown("### Table 5: Open-Loop Planning Results on nuScenes")
            gr.Markdown("**Metrics:** L2 Error (m) ↓ | Collision Rate ↓")
            nuscenes_stats = gr.Markdown(initial_nuscenes_stats)
            with gr.Row():
                nuscenes_category = gr.Dropdown(label="Category", choices=["All", "Vision-Action", "Vision-Language-Action"], value="All")
                nuscenes_search = gr.Textbox(label="Search Model", placeholder="e.g., UniAD, EMMA...")
            nuscenes_update_btn = gr.Button("🔍 Update", variant="primary")
            nuscenes_table = gr.Dataframe(value=initial_nuscenes_df, label="nuScenes Results", interactive=False, wrap=True)
            nuscenes_plot = gr.Plot(label="Statistics", value=initial_nuscenes_plot, format="png")
            nuscenes_update_btn.click(fn=lambda cat, search: (*update_nuscenes_table(cat, search), create_nuscenes_plot(cat)),
                                     inputs=[nuscenes_category, nuscenes_search], 
                                     outputs=[nuscenes_table, nuscenes_stats, nuscenes_plot])
        
        # Tab 5: WOD-E2E
        with gr.Tab("WOD-E2E"):
            gr.Markdown("### Table 6: Results on WOD-E2E Test Split")
            gr.Markdown("**Metrics:** RFS (Overall/Spotlight) ↑ | ADE (5s/3s) ↓")
            wod_stats = gr.Markdown(initial_wod_stats)
            with gr.Row():
                wod_category = gr.Dropdown(label="Category", choices=["All", "Vision-Action", "Vision-Language-Action"], value="All")
                wod_search = gr.Textbox(label="Search Model", placeholder="e.g., AutoVLA...")
            wod_update_btn = gr.Button("🔍 Update", variant="primary")
            wod_table = gr.Dataframe(value=initial_wod_df, label="WOD-E2E Results", interactive=False, wrap=True)
            wod_plot = gr.Plot(label="Statistics", value=initial_wod_plot, format="png")
            wod_update_btn.click(fn=lambda cat, search: (*update_wod_table(cat, search), create_wod_plot(cat)),
                                inputs=[wod_category, wod_search], 
                                outputs=[wod_table, wod_stats, wod_plot])
        
        # Tab 6: NAVSIM
        with gr.Tab("NAVSIM Closed-Loop"):
            gr.Markdown("### Table 7: Closed-Loop Results on NAVSIM")
            gr.Markdown("**Metrics:** NC/DAC/TTC/Comf/EP/PDMS ↑")
            navsim_stats = gr.Markdown(initial_navsim_stats)
            with gr.Row():
                navsim_category = gr.Dropdown(label="Category", choices=["All", "Vision-Action", "Vision-Language-Action"], value="All")
                navsim_search = gr.Textbox(label="Search Model", placeholder="e.g., ReflectDrive...")
            navsim_update_btn = gr.Button("🔍 Update", variant="primary")
            navsim_table = gr.Dataframe(value=initial_navsim_df, label="NAVSIM Results", interactive=False, wrap=True)
            navsim_plot = gr.Plot(label="Statistics", value=initial_navsim_plot, format="png")
            navsim_update_btn.click(fn=lambda cat, search: (*update_navsim_table(cat, search), create_navsim_plot(cat)),
                                   inputs=[navsim_category, navsim_search], 
                                   outputs=[navsim_table, navsim_stats, navsim_plot])
        
        # Tab 7: Bench2Drive
        with gr.Tab("Bench2Drive"):
            gr.Markdown("### Table 8: Closed-Loop & Open-Loop Results on Bench2Drive")
            gr.Markdown("**Metrics:** DS/SR ↑ | L2 Avg ↓")
            bench2drive_stats = gr.Markdown(initial_bench2drive_stats)
            with gr.Row():
                bench2drive_category = gr.Dropdown(label="Category", choices=["All", "Vision-Action", "Vision-Language-Action"], value="All")
                bench2drive_search = gr.Textbox(label="Search Model", placeholder="e.g., SimLingo...")
            bench2drive_update_btn = gr.Button("🔍 Update", variant="primary")
            bench2drive_table = gr.Dataframe(value=initial_bench2drive_df, label="Bench2Drive Results", interactive=False, wrap=True)
            bench2drive_plot = gr.Plot(label="Statistics", value=initial_bench2drive_plot, format="png")
            bench2drive_update_btn.click(fn=lambda cat, search: (*update_bench2drive_table(cat, search), create_bench2drive_plot(cat)),
                                        inputs=[bench2drive_category, bench2drive_search], 
                                        outputs=[bench2drive_table, bench2drive_stats, bench2drive_plot])
    
    gr.Markdown("""
    ---
    **Legend:** 📷 Camera | 🔍 LiDAR | ⚙️ Status | 💭 Prompt | 📝 Instruction | 🌆 Scene | 🚦 Traffic | 📚 Context
    
    **Action Types:** RL=Reinforcement Learning | REG=Decoder+MLP | SEL=Selection | GEN=Generative | LH=Language Head
    
    📖 **Project:** [https://worldbench.github.io/vla4ad](https://worldbench.github.io/vla4ad) | 🔗 **GitHub:** [https://github.com/worldbench/awesome-vla-for-ad](https://github.com/worldbench/awesome-vla-for-ad)
    """)

if __name__ == "__main__":
    demo.launch()