v0.1.17; update 7-9

2025-12-27 16:13:18 +00:00
parent f92f34ac14
commit baf11a6cd4
12 changed files with 405 additions and 114 deletions
--- a/.gitignore
+++ b/.gitignore
@@ -8,5 +8,6 @@ __pycache__/

 # Logger
 logs/
+logs*/
 .aim/
 *.bkp
--- a/CMD.md
+++ b/CMD.md
@@ -51,9 +51,8 @@ python robogauge/scripts/run.py \
    --task go2_moe \
    --experiment-name debug \
    --stress-benchmark \
-    --stress-terrain-names flat slope stairs_up stairs_down wave \
-    --stress-num-processes 2 \
-    --num-processes 3 \
+    --stress-terrain-names flat slope stairs_up stairs_down wave obstacle \
+    --num-processes 6 \
    --seeds 0 1 2 3 4 \
    --frictions 0.5 1.0 1.5 2.0 2.5 \
    --compress-logs \
@@ -64,14 +63,23 @@ python robogauge/scripts/run.py \
 将Multi Run结果绘制在雷达图中
 ```bash
 # 可选--out保存到图片
-python robogauge/utils/visualize_results.py \
+python robogauge/utils/visualize/plot_radar_and_bar.py \
    aggregated_results_1.yaml \
    aggregated_results_2.yaml \
    --out logs/1.jpg

 # 绘制下全部 *.yaml, 可选range控制radar, bar y轴显示范围
-python robogauge/utils/visualize_results.py \
+python robogauge/utils/visualize/plot_radar_and_bar.py \
    /home/xfy/Coding/robot_gauge/mytest/results \
    --range 0.35 1.0 \
    --out logs/1.jpg
-```
+```
+
+# Terrain Levels Plot
+```bash
+# 绘制5种可变等级的地形与摩擦系数的关系图
+python robogauge/utils/visualize/plot_terrain_levels.py \
+    aggregated_results_1.yaml \
+    aggregated_results_2.yaml \
+    --out logs/terrain_analysis.jpg
+```
--- a/README.md
+++ b/README.md
@@ -13,10 +13,11 @@
 | Unitree go2 | 四足 | ✅ |

 ## 安装
-参考[PyTorch官网](https://pytorch.org/get-started/locally/)安装2.0以上任意版本, 安装本仓库
+推荐安装`python=3.8`, 参考[PyTorch官网](https://pytorch.org/get-started/locally/)安装2.0以上任意版本, 安装本仓库

 ```bash
 git clone https://github.com/wty-yy/RoboGauge.git
+cd RoboGauge
 pip install -e .
 ```

--- a/UPDATE.md
+++ b/UPDATE.md
@@ -6,7 +6,10 @@
 3. 在StressPipeline测试中, 不再保存
 4. 添加obstacles地形
 5. 优化分数计算方法, 按照`(level-1)*0.1+metric_score`计算
-6. 加入日志文件夹递归保存, StressPipeline -> LevelPipeline -> MultiPipeline, 加入压缩参数`--compress-logs`, 将日志文件夹压缩为`subtasks.tar.xz`, 日志文件压缩`4M->100kb`
+6. 加入日志文件夹递归保存, StressPipeline -> LevelPipeline -> MultiPipeline, 加入压缩参数`--compress-logs`, 将日志文件夹压缩为`subtasks.tar.xz`, 日志文件压缩`4M->100kb` (6种地形的全部评估测试日志压缩后`29M->539K`)
+7. StressPipeline和MultiPipeline同时工作时, 很大概率发生死锁 (StressPipeline传入的任务超过10个以上), 将MultiPipeline的num-processes设置为1即可避免该问题 (不创建Process), 为了提高速度可以将flat的不同域随机化也展平
+8. 将进度条换成当前时间/剩余时间
+9. 添加`plot_terrain_levels.py`绘图代码, 绘制5种可变等级的地形与摩擦系数的关系图
 ## 20251226
 ### v0.1.16
 1. 基本完成StressPipeline, 加入绘制进度条的线程, 其他进程通过Queue更新主进程的进度条
--- a/robogauge/tasks/pipeline/multi_pipeline.py
+++ b/robogauge/tasks/pipeline/multi_pipeline.py
@@ -91,25 +91,38 @@ class MultiPipeline:
        multi_logger.info(f"🔢 Seeds: {self.seeds}, Frictions: {self.frictions}, Base masses: {self.base_masses}")

        workers_data = list(product(self.seeds, self.base_masses, self.frictions))
-        report_progress(self.progress_data, ProgressTypes.INIT, total=len(workers_data), desc="🚀 MultiPipeline Executing")
+        report_progress(self.progress_data, ProgressTypes.INIT, total=len(workers_data), desc="🚀 MultiPipeline")

        ctx = multiprocessing.get_context('spawn')
        worker_func = functools.partial(run_single_process, self.args)
        results_list = []
-        with NoDaemonPool(processes=self.num_processes, context=ctx) as pool:
-            iterator = pool.imap_unordered(worker_func, workers_data)
-            bar = iterator
+
+        def update_results(results):
+            results_list.append(results)
+            self.add_static_info('model_path', results['model_path'])
+            self.add_static_info('terrain_name', results['results']['terrain_name'])
+            self.add_static_info('terrain_level', results['results']['terrain_level'])
+            report_progress(self.progress_data, ProgressTypes.UPDATE, value=1)
+            if results['status'] != 'success':
+                data = results['data']
+                multi_logger.error(f"❌ Process with seed={data[0]}, base_mass={data[1]}, friction={data[2]} failed with error: {results['error_msg']}")
+
+        if self.num_processes == 1:
+            multi_logger.info("🚀 Running in Serial Mode")
+            bar = workers_data
            if self.console_output:
-                bar = tqdm(iterator, total=len(workers_data), desc="Evaluation")
-            for results in bar:
-                results_list.append(results)
-                self.add_static_info('model_path', results['model_path'])
-                self.add_static_info('terrain_name', results['results']['terrain_name'])
-                self.add_static_info('terrain_level', results['results']['terrain_level'])
-                report_progress(self.progress_data, ProgressTypes.UPDATE, value=1)
-                if results['status'] != 'success':
-                    data = results['data']
-                    multi_logger.error(f"❌ Process with seed={data[0]}, base_mass={data[1]}, friction={data[2]} failed with error: {results['error_msg']}")
+                bar = tqdm(workers_data, desc="Evaluation")
+            for data in bar:
+                results = worker_func(data)
+                update_results(results)
+        else:
+            with NoDaemonPool(processes=self.num_processes, context=ctx) as pool:
+                iterator = pool.imap_unordered(worker_func, workers_data)
+                bar = iterator
+                if self.console_output:
+                    bar = tqdm(iterator, total=len(workers_data), desc="Evaluation")
+                for results in bar:
+                    update_results(results)

        multi_logger.info("✅ Multi-Process Evaluation Completed.")
        aggregated_results = self.aggregate_results(results_list)
--- a/robogauge/tasks/pipeline/stress_pipeline.py
+++ b/robogauge/tasks/pipeline/stress_pipeline.py
@@ -41,23 +41,22 @@ def run_pipeline(args, progress_queue, data):
    args = deepcopy(args)
    task_id = data['task_id']
    search = data['search_max_level']
-    task_label = f"[{data['terrain_name']}]"
+    task_label = f"[{data['terrain_name']}] M:{data['base_mass']} F:{data['friction']}"
    progress_data = ProgressData(
        task_id=task_id,
        msg_prefix=task_label + ' ',
        progress_queue=progress_queue
    )
-    if search is True:
-        task_label += f" M:{data['base_mass']} F:{data['friction']}"
-        progress_data.msg_prefix = task_label + ' '
-        args.friction = data['friction']
-        args.frictions = [data['friction']]
-        args.base_mass = data['base_mass']
-        args.base_masses = [data['base_mass']]
-        args.task_name = f"{data['task_robot_model']}.{data['terrain_name']}"
-        args.experiment_name = f"{args.experiment_name}_{data['terrain_name']}_M{data['base_mass']}_F{data['friction']}"
-        args.goals = GOALS['level_pipeline']

+    args.friction = data['friction']
+    args.frictions = [data['friction']]
+    args.base_mass = data['base_mass']
+    args.base_masses = [data['base_mass']]
+    args.task_name = f"{data['task_robot_model']}.{data['terrain_name']}"
+    args.experiment_name = f"{args.experiment_name}_{data['terrain_name']}_M{data['base_mass']}_F{data['friction']}"
+
+    if search is True:
+        args.goals = GOALS['level_pipeline']
        level, level_results = LevelPipeline(args, console_output=False, progress_data=progress_data).run()
        if level == 0:  # no valid level found
            report_progress(progress_data, ProgressTypes.FINISH, desc=f"❌ Failed (Lv 0)")
@@ -68,8 +67,8 @@ def run_pipeline(args, progress_queue, data):
                'level': 0,
            }
            return results
-        
        report_progress(progress_data, ProgressTypes.RESET, total=0, desc=f"✅ Found Lv {level} -> Running")
+        progress_data.msg_prefix += f"(Lv {level}) "
    else:
        level = None  # flat terrain
        args.task_name = f"{data['task_robot_model']}.{data['terrain_name']}"
@@ -91,13 +90,14 @@ class StressPipeline:
        self.args = args
        self.seeds = args.seeds
        self.task_robot_model = args.task_name.split('.')[0]
-        self.num_processes = args.stress_num_processes
+        self.num_processes = args.num_processes
        args.experiment_name = self.task_robot_model + '_stress' + ('' if args.cli_experiment_name is None else '_' + args.cli_experiment_name)
        self.static_info = {}
        stress_logger.create(args.experiment_name, args.run_name)
        self.args.parent_log_dir = str(stress_logger.log_dir / "subtasks")
        self.compress_logs = args.compress_logs
        args.compress_logs = False  # Disable child log compression
+        args.num_processes = 1  # Disable child multi-process

    def add_static_info(self, key: str, value):
        if key not in self.static_info:
@@ -125,16 +125,13 @@ class StressPipeline:
                'terrain_name': terrain_name,
                'search_max_level': search_max_level,
            }
-            if search_max_level:
-                for friction, base_mass in product(self.args.frictions, self.args.base_masses):
-                    now_data = deepcopy(data)
-                    now_data.update({
-                        'friction': friction,
-                        'base_mass': base_mass,
-                    })
-                    workers_data.append(now_data)
-            else:
-                workers_data.append(data)
+            for friction, base_mass in product(self.args.frictions, self.args.base_masses):
+                now_data = deepcopy(data)
+                now_data.update({
+                    'friction': friction,
+                    'base_mass': base_mass,
+                })
+                workers_data.append(now_data)

        ### Start progress monitor ###
        progress_queue, monitor_thread = start_progress_monitor_thread(len(workers_data))
@@ -182,16 +179,14 @@ class StressPipeline:
            stress_logger.error("No results to aggregate.")
            return

-        summary = {**self.static_info, 'summary': {}}
+        summary = {**self.static_info, 'summary': {}, 'final_score': {}}
        value_collections = defaultdict(lambda: defaultdict(list))
        zero_terrain_count = 0
        for result in all_results:
            terrain_name = result['data']['terrain_name']
            terrain_level = result['level']  # None, 0, 1, ..., 10
-            key = terrain_name
-            if terrain_level is not None:
-                key += f'_{terrain_level}'
-                key += f'_baseMass{result["data"]["base_mass"]}_friction{result["data"]["friction"]}'
+            key = f'{terrain_name}_{terrain_level}'
+            key += f'_baseMass{result["data"]["base_mass"]}_friction{result["data"]["friction"]}'
            if terrain_level == 0:
                summary[key] = None
                zero_terrain_count += 1
@@ -201,14 +196,19 @@ class StressPipeline:
            for metric, means in result['results']['summary'].items():
                for mean_name, mean_value in means.items():
                    value = float(mean_value.split(' ')[0])
-                    value += (terrain_level - 1) * 0.1 if terrain_level is not None else 0.0
+                    if terrain_level is not None:  # level terrain
+                        value = (terrain_level - 1) * 0.1 + value * 0.1
                    value_collections[metric][mean_name].append(value)

+        final_metrics = defaultdict(list)
        for metric, means in value_collections.items():
            summary['summary'][metric] = {}
            for mean_name, values in means.items():
                values.extend([0.0] * zero_terrain_count)  # include zero terrains
                summary['summary'][metric][mean_name] = f"{float(np.mean(values)):.4f} ± {float(np.std(values)):.4f}"
+                final_metrics[mean_name].append(np.mean(values))
+        for mean_name, values in final_metrics.items():
+            summary['final_score'][mean_name] = float(np.mean(values))

        save_path = stress_logger.log_dir / "stress_benchmark_results.yaml"
        with open(save_path, 'w') as file:
--- a/robogauge/utils/helpers.py
+++ b/robogauge/utils/helpers.py
@@ -80,7 +80,6 @@ def parse_args():

        # Multiprocessing parameters, with different seeds
        {"name": "--multi", "action": "store_true", "default": False, "help": "Enable multiprocessing."},
-        {"name": "--num-processes", "type": int, "default": 2, "help": "Number of parallel processes."},
        {"name": "--seeds", "type": int, "nargs": "+", "default": [0, 1, 2, 3, 4], "help": "List of random seeds for multiple runs."},
        {"name": "--base-masses", "type": float, "nargs": "+", "default": [0], "help": "List of base masses for the model."},
        {"name": "--frictions", "type": float, "nargs": "+", "default": [0.5, 0.75, 1.0, 1.25, 1.5, 1.75, 2.0, 2.25, 2.5], "help": "List of friction coefficients for the model."},
@@ -91,8 +90,9 @@ def parse_args():
        # Stress pipeline parameters
        {"name": "--stress-benchmark", "action": "store_true", "default": False, "help": "Use stress pipeline to benchmark model robustness."},
        {"name": "--stress-terrain-names", "type": str, "nargs": "+", "default": ["flat", "slope", "wave", "stairs_up", "stairs_down"], "help": "List of terrain names for stress benchmark."},
-        {"name": "--stress-num-processes", "type": int, "default": 2, "help": "Number of parallel processes for stress benchmark."},

+        # Common parameters
+        {"name": "--num-processes", "type": int, "default": 2, "help": "Number of parallel processes for Multi or Stress benchmark."},
        {"name": "--compress-logs", "action": "store_true", "default": False, "help": "Compress and delete logs after run."},
    ]
    for param in parameters:
--- a/robogauge/utils/progress_monitor.py
+++ b/robogauge/utils/progress_monitor.py
@@ -2,16 +2,17 @@
 '''
@File    : progress_monitor.py
@Time    : 2025/12/27 00:10:07
-@Author  : wty-yy
+@Author  : wty-yy (with Gemini 3)
@Version : 1.0
@Blog    : https://wty-yy.github.io/
-@Desc    : Centralized progress monitoring for multiprocessing tasks using tqdm
+@Desc    : Centralized progress monitoring with Dynamic Slot Management
 '''
 from tqdm import tqdm
 import multiprocessing
 from typing import Dict
 from threading import Thread
 from dataclasses import dataclass
+import heapq

 class ProgressTypes:
    INIT = 'init'       # Init progress (set total, desc)
@@ -50,76 +51,111 @@ def report_progress(progress_data: ProgressData, msg_type, value=None, desc=None
 class ProgressMonitor:
    def __init__(self, total_rows):
        self.total_rows = total_rows
-        self.bars: Dict[int, tqdm] = {}
+        self.active_bars: Dict[int, Dict] = {} 
+        self.free_slots = []
+        self.max_slot_used = 0 
        
+    def _get_free_slot(self):
+        """ Get a free screen display position """
+        if self.free_slots:
+            return heapq.heappop(self.free_slots)
+        else:
+            self.max_slot_used += 1
+            return self.max_slot_used
+
+    def _release_slot(self, slot):
+        """ Release a screen display position """
+        heapq.heappush(self.free_slots, slot)
+
    def listener_loop(self, queue):
        """
-        Run in a separate thread in the main process to consume the Queue and update tqdm.
+        Listener thread in the main process
        """
-        # print(f"Monitor started for {self.total_rows} tasks...")
-        for i in range(self.total_rows):
-            self.bars[i] = tqdm(
-                total=100, 
-                position=i, 
-                desc=f"Task {i} Pending...", 
-                bar_format="{l_bar}{bar}| {n_fmt}/{total_fmt} [{elapsed}]",
-                leave=True
-            )
+        # Create the main progress bar (Position 0)
+        main_bar = tqdm(
+            total=self.total_rows, 
+            position=0, 
+            desc="🚀 Total Progress", 
+            bar_format="{l_bar}{bar}| {n_fmt}/{total_fmt} [Done: {n_fmt}] [{elapsed}<{remaining}]",
+            leave=True,
+            smoothing=0.1  # Optional: set smoothing factor (0-1) to prevent drastic ETA fluctuations due to slow tasks
+        )
        
-        active_tasks = self.total_rows
+        # Number of tasks still pending
+        pending_tasks = self.total_rows
        
-        while active_tasks > 0:
+        while pending_tasks > 0:
            record = queue.get()
-            if record is None:  # Poison pill signal
-                break
+            if record is None: break
            
            task_id, msg_type, data = record
-            if task_id not in self.bars:
-                continue
-                
-            bar = self.bars[task_id]
            
-            if msg_type == ProgressTypes.INIT:
-                total = data['total']
-                desc = data['desc']
-                bar.reset(total=total)
-                bar.set_description(desc)
-                bar.refresh()
-                
-            elif msg_type == ProgressTypes.UPDATE:
-                val = data['value']
-                bar.update(val)
-                
-            elif msg_type == ProgressTypes.DESC:
-                desc = data['desc']
-                if desc:
+            # --- Dynamically create/get sub progress bars ---
+            if task_id not in self.active_bars:
+                # Only create bar when message received to avoid Pending screen flicker
+                slot = self._get_free_slot()
+                # leave=False is key: after task completion, clear the line for new task reuse
+                bar = tqdm(
+                    total=100,
+                    position=slot,
+                    desc=f"Task {task_id} Starting...",
+                    bar_format="{l_bar}{bar}| {n_fmt}/{total_fmt} [{elapsed}<{remaining}]",
+                    leave=False 
+                )
+                self.active_bars[task_id] = {'bar': bar, 'slot': slot}
+            
+            bar_info = self.active_bars[task_id]
+            bar = bar_info['bar']
+            
+            # --- Handle messages ---
+            try:
+                if msg_type == ProgressTypes.INIT:
+                    total = data.get('total', 100)
+                    desc = data.get('desc', '')
+                    bar.reset(total=total)
                    bar.set_description(desc)
+                    bar.refresh()
                    
-            elif msg_type == ProgressTypes.RESET:
-                # Scenario: Level search finished, starting MultiPipeline, reset progress bar
-                total = data['total']
-                desc = data['desc']
-                bar.reset(total=total)
-                bar.set_description(desc)
-                bar.refresh()
-                
-            elif msg_type == ProgressTypes.FINISH:
-                desc = data['desc']
-                if desc:
+                elif msg_type == ProgressTypes.UPDATE:
+                    val = data.get('value', 1)
+                    bar.update(val)
+                    
+                elif msg_type == ProgressTypes.DESC:
+                    desc = data.get('desc')
+                    if desc: bar.set_description(desc)
+                        
+                elif msg_type == ProgressTypes.RESET:
+                    total = data.get('total', 100)
+                    desc = data.get('desc', '')
+                    bar.reset(total=total)
                    bar.set_description(desc)
-                bar.refresh()
-                # Note: We do not close the bar here to keep it displayed until all tasks are finished and closed together.
-                active_tasks -= 1
-                
-            elif msg_type == ProgressTypes.ERROR:
-                desc = data['desc']
-                bar.set_description(desc)
-                bar.refresh()
-                active_tasks -= 1
+                    bar.refresh()
+                    
+                elif msg_type == ProgressTypes.FINISH or msg_type == ProgressTypes.ERROR:
+                    # Update sub progress bar status
+                    desc = data.get('desc')
+                    if desc: bar.set_description(desc)
+                    bar.refresh()
+                    
+                    # Close sub progress bar and release slot
+                    bar.close() # 因为 leave=False，这行会从屏幕消失
+                    slot = bar_info['slot']
+                    self._release_slot(slot)
+                    del self.active_bars[task_id]
+                    
+                    # Update the main progress bar
+                    main_bar.update(1)
+                    main_bar.refresh()
+                    
+                    pending_tasks -= 1

-        # After all tasks are finished, close all bars
-        for bar in self.bars.values():
-            bar.close()
+            except Exception as e:
+                pass # Prevent printing errors from disrupting layout
+
+        # Close all remaining bars (theoretically should be empty except main_bar)
+        main_bar.close()
+        for info in self.active_bars.values():
+            info['bar'].close()

 def start_progress_monitor_thread(total_rows):
    """
--- a/robogauge/utils/visualize/plot_radar_and_bar.py
+++ b/robogauge/utils/visualize/plot_radar_and_bar.py
@@ -1,3 +1,12 @@
+# -*- coding: utf-8 -*-
+'''
+@File    : plot_radar_and_bar.py
+@Time    : 2025/12/27 22:52:24
+@Author  : wty-yy (with Gemini 3)
+@Version : 1.0
+@Blog    : https://wty-yy.github.io/
+@Desc    : 可视化评测结果，生成雷达图和柱状图
+'''
 import matplotlib.pyplot as plt
 from pathlib import Path
 import numpy as np
@@ -48,9 +57,10 @@ def load_data(file_paths):
        model_name = os.path.basename(raw_path).replace('.pt', '')

        values = []
+        summary = content['summary']
        for k in metric_keys:
-            if k in content:
-                raw_val = content[k]['mean@50']
+            if k in summary:
+                raw_val = summary[k]['mean@25']
                values.append(parse_value_string(raw_val))
            else:
                values.append(0.0)
--- a/robogauge/utils/visualize/plot_terrain_levels.py
+++ b/robogauge/utils/visualize/plot_terrain_levels.py
@@ -0,0 +1,168 @@
+# -*- coding: utf-8 -*-
+'''
+@File    : plot_terrain_levels.py
+@Time    : 2025/12/27 23:04:01
+@Author  : wty-yy (with Gemini 3)
+@Version : 1.0
+@Blog    : https://wty-yy.github.io/
+@Desc    : None
+@Desc    : 可视化不同地形下随着摩擦力变化的关卡通过等级 (Terrain Level vs Friction)
+'''
+import matplotlib.pyplot as plt
+from pathlib import Path
+import numpy as np
+import yaml
+import argparse
+import os
+import re
+
+# --- 配置 Matplotlib 样式 ---
+config = {
+    "font.family": 'serif',
+    "figure.figsize": (18, 10),
+    "font.size": 12,
+    "mathtext.fontset": 'cm',
+    'axes.unicode_minus': False
+}
+plt.rcParams.update(config)
+plt.rcParams['font.serif'] = ['Times New Roman', 'DejaVu Serif', 'serif']
+
+def load_terrain_data(file_paths):
+    """
+    加载数据并按模型、地形、摩擦力组织
+    Returns:
+        model_data: {model_name: {terrain_type: {friction: level}}}
+    """
+    target_terrains = ['wave', 'slope', 'stairs_up', 'stairs_down', 'obstacle']
+    model_data = {}
+    
+    # 用于从键名中提取 friction 的正则 (例如 friction1.25)
+    friction_pattern = re.compile(r'friction([\d\.]+)')
+
+    for path in file_paths:
+        if not os.path.exists(path):
+            print(f"[ERROR] File not found: {path}")
+            continue
+            
+        with open(path, 'r', encoding='utf-8') as f:
+            content = yaml.safe_load(f)
+        
+        # 获取模型名称
+        raw_path = content.get('model_path', 'Unknown_Model')
+        model_name = os.path.basename(raw_path).replace('.pt', '')
+        
+        if model_name not in model_data:
+            model_data[model_name] = {t: {} for t in target_terrains}
+            
+        # 遍历YAML中的每个测试条目
+        for key, value in content.items():
+            if key in ['model_path', 'summary']:
+                continue
+            
+            # 1. 检查地形是否在目标列表中
+            t_name = value.get('terrain_name')
+            if t_name not in target_terrains:
+                continue
+            
+            # 2. 从 key 中提取摩擦力
+            match = friction_pattern.search(key)
+            if not match:
+                continue
+            friction = float(match.group(1))
+            
+            # 3. 获取地形等级 (取该条件下该地形的最大值，防止重复)
+            level = value.get('terrain_level', 0)
+            
+            # 记录数据
+            current_max = model_data[model_name][t_name].get(friction, -1)
+            if level > current_max:
+                model_data[model_name][t_name][friction] = level
+                
+    return model_data, target_terrains
+
+def plot_results(model_data, terrain_labels, output_file=None):
+    """绘制 2x3 的子图"""
+    if not model_data:
+        print("No data to plot.")
+        return
+
+    # 设置 2行3列 的布局
+    fig, axes = plt.subplots(2, 3, figsize=(18, 10), sharex=True, sharey=True)
+    axes = axes.flatten()
+    
+    # 自动获取颜色映射
+    colors = plt.cm.get_cmap("tab10", len(model_data))
+    
+    # 遍历前5个地形进行绘图
+    for i, t_name in enumerate(terrain_labels):
+        ax = axes[i]
+        
+        for idx, (model_name, terrains) in enumerate(model_data.items()):
+            data_points = terrains.get(t_name, {})
+            if not data_points:
+                continue
+            
+            # 排序 (Friction, Level)
+            sorted_points = sorted(data_points.items())
+            xs, ys = zip(*sorted_points)
+            
+            ax.plot(xs, ys, marker='o', linestyle='-', linewidth=2, 
+                    label=model_name, color=colors(idx), alpha=0.8)
+        
+        ax.set_title(t_name.replace('_', ' ').title(), fontsize=14)
+        ax.grid(True, linestyle='--', alpha=0.6)
+
+    # --- 第6张图：绘制平均值 (Average) ---
+    ax_avg = axes[5]
+    for idx, (model_name, terrains) in enumerate(model_data.items()):
+        # 聚合该模型下所有地形的数据用于计算平均值
+        # 结构: friction -> [level_slope, level_wave, ...]
+        friction_agg = {}
+        for t_name in terrain_labels:
+            for f, l in terrains[t_name].items():
+                if f not in friction_agg:
+                    friction_agg[f] = []
+                friction_agg[f].append(l)
+        
+        if not friction_agg:
+            continue
+            
+        # 计算平均值并排序
+        sorted_frictions = sorted(friction_agg.keys())
+        avg_levels = [np.mean(friction_agg[f]) for f in sorted_frictions]
+        
+        ax_avg.plot(sorted_frictions, avg_levels, marker='s', linestyle='-', linewidth=2,
+                    label=model_name, color=colors(idx), alpha=0.8)
+
+    ax_avg.set_title('Average Performance', fontsize=14)
+    ax_avg.grid(True, linestyle='--', alpha=0.6)
+    
+    # 仅在最后一张图显示图例，避免遮挡
+    ax_avg.legend(loc='best', frameon=False, title="Models")
+
+    # 最下面一行设置 X 轴标签
+    for i in [3, 4, 5]:
+        axes[i].set_xlabel('Friction ($\mu$)')
+
+    # 最左边一列设置 Y 轴标签
+    for i in [0, 3]:
+        axes[i].set_ylabel('Terrain Level')
+
+    plt.tight_layout()
+    # 稍微调整间距防止标签重叠
+    # plt.subplots_adjust(wspace=0.1, hspace=0.15)
+    
+    if output_file:
+        plt.savefig(output_file, dpi=300, bbox_inches='tight')
+        print(f"Plot saved to {output_file}")
+    else:
+        plt.show()
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Plot terrain level vs friction curves.")
+    parser.add_argument('files', metavar='F', type=str, nargs='+', help='YAML result files')
+    parser.add_argument('--out', type=str, default='terrain_analysis.jpg', help='Output image filename')
+    args = parser.parse_args()
+    
+    data, labels = load_terrain_data(args.files)
+    plot_results(data, labels, output_file=args.out)
--- a/run.bash
+++ b/run.bash
@@ -0,0 +1,49 @@
+#!/bin/bash
+
+source /root/Programs/miniforge3/bin/activate robot
+
+python robogauge/scripts/run.py \
+    --task go2_moe \
+    --experiment-name debug \
+    --stress-benchmark \
+    --stress-terrain-names flat slope stairs_up stairs_down wave obstacle \
+    --num-processes 62 \
+    --seeds 0 1 2 3 4 \
+    --frictions 0.75 1.0 1.25 1.5 1.75 2.0 2.25 2.5 \
+    --compress-logs \
+    --headless
+
+python robogauge/scripts/run.py \
+    --task go2_moe \
+    --model-path /root/Coding/RoboGauge/mytest/go2_moe_cts_hard_terrain_141k.pt \
+    --experiment-name debug \
+    --stress-benchmark \
+    --stress-terrain-names flat slope stairs_up stairs_down wave obstacle \
+    --num-processes 62 \
+    --seeds 0 1 2 3 4 \
+    --frictions 0.75 1.0 1.25 1.5 1.75 2.0 2.25 2.5 \
+    --compress-logs \
+    --headless
+
+python robogauge/scripts/run.py \
+    --task go2 \
+    --model-path /root/Coding/RoboGauge/mytest/go2_cts_hard_terrain_141k.pt \
+    --experiment-name debug \
+    --stress-benchmark \
+    --stress-terrain-names flat slope stairs_up stairs_down wave obstacle \
+    --num-processes 62 \
+    --seeds 0 1 2 3 4 \
+    --frictions 0.75 1.0 1.25 1.5 1.75 2.0 2.25 2.5 \
+    --compress-logs \
+    --headless
+
+python robogauge/scripts/run.py \
+    --task go2 \
+    --experiment-name debug \
+    --stress-benchmark \
+    --stress-terrain-names flat slope stairs_up stairs_down wave obstacle \
+    --num-processes 62 \
+    --seeds 0 1 2 3 4 \
+    --frictions 0.75 1.0 1.25 1.5 1.75 2.0 2.25 2.5 \
+    --compress-logs \
+    --headless
--- a/setup.py
+++ b/setup.py
@@ -13,9 +13,11 @@ setup(
        "mujoco>=3.0.0",
        "dm_control>=1.0.14",
        "scipy",
-        "matplotlib",
+        "matplotlib==3.6.3",
        "tqdm",
        "imageio[ffmpeg]",
+        "tensorboard",
+        "PyYAML",
    ],
    python_requires=">=3.8",