diff --git a/docs/humanoid/g1_costnet_grid_planner/index.md b/docs/humanoid/g1_costnet_grid_planner/index.md
new file mode 100644
index 0000000000..c4d14643e7
--- /dev/null
+++ b/docs/humanoid/g1_costnet_grid_planner/index.md
@@ -0,0 +1,48 @@
+# G1 CostNet 栅格规划对照
+
+在独立栅格示例中比较 CostNet 辅助 A* 和曼哈顿 A* 的路径代价与搜索扩展数。
+
+## 环境与运行
+
+使用 Python 3.11。在模块目录运行 `python -m pip install -r requirements.txt`。
+
+```bash
+python main.py --output results
+```
+
+示例在一个合成栅格比较经典和神经辅助搜索,输出路线图和 JSON。网络仅处理相同 f 值的排序。可用 `--mode train_planner` 重训,用 `--mode plan_stdio` 接收占据栅格 JSON。
+
+## ROS 1 Noetic 启动
+
+本模块的 ROS 接口使用 ROS 1 Noetic(catkin、rospy)。在 Noetic 源码空间运行,将路径替换为实际值。
+
+```bash
+source /opt/ros/noetic/setup.bash
+export ROS_PACKAGE_PATH="$PWD:${ROS_PACKAGE_PATH:-}"
+roslaunch g1_costnet_grid_planner main.launch sim_python:=/path/to/python sim_entry:=/path/to/main.py
+```
+
+## 运行效果
+
+
+
+## 可复核结果
+
+| 方法或指标 | 路径代价(格) | 扩展节点 |
+| --- | ---: | ---: |
+| Manhattan A* | 48 | 401 |
+| 随仓库提供的 CostNet 辅助 A* | 48 | 49 |
+| 40 张合成测试图的 CostNet MAE | 0.756 | — |
+
+扩展数仅对应演示地图。推理按栅格各轴尺寸归一化相对坐标,各轴尺度至少为 32,以保持 32×32 训练输入及已有权重兼容。大图的搜索扩展数会随布局变化。
+
+```bash
+python main.py --mode train_planner --epochs 5 --samples 100 --seed 91 --weights results/trained
+python main.py --mode plan_stdio --robot-radius 0.3 --stride 3 < request.json
+```
+
+请求 JSON 包含 `map`(行优先的一维占据数组)、`width`、`height`、`resolution`(米/格)、`origin`、`start`、`goal`(二维世界坐标)。可用 `robot_radius`、`stride` 字段覆盖命令行默认值。重复在同一 Python 进程调用时复用权重,文件更新后重新加载。
+
+## 大模型使用声明
+
+使用 OpenAI Codex 辅助拆分、编程、验证和文档整理,提交者复核并对提交内容负责。
diff --git a/docs/img/g1_costnet_grid_planner/planner.png b/docs/img/g1_costnet_grid_planner/planner.png
new file mode 100644
index 0000000000..36c4c600f4
Binary files /dev/null and b/docs/img/g1_costnet_grid_planner/planner.png differ
diff --git a/docs/index.md b/docs/index.md
index 9f6f16ff53..4551cacf5a 100644
--- a/docs/index.md
+++ b/docs/index.md
@@ -99,6 +99,7 @@ ROS 2 服务与参数基础实验(无人机模式切换与在线调参):[c
* [人形上下台阶步态仿真](./humanoid/humanoid_stair_climb.md)
* [ROS虚拟机器人圆形轨迹rviz可视化示例](./humanoid/ros_topic_demo/index.md)
* [人形机器人全身逆运动学与可操作性](./humanoid/whole_body_ik.md)
+* [G1 CostNet 栅格规划对照](./humanoid/g1_costnet_grid_planner/index.md)
* [人形机器人双臂末端可达域分析](./humanoid/humanoid_reachability.md)
* [人形机器人跌倒预判与保护性倒地](./humanoid/humanoid_fall_protection.md)
* [人形机器人 3D 碰撞体边界检测可视化](./humanoid/humanoid_collision_boundary.md)
diff --git a/mkdocs.yml b/mkdocs.yml
index edf15e5906..dabd97e228 100644
--- a/mkdocs.yml
+++ b/mkdocs.yml
@@ -97,6 +97,7 @@ nav:
- ROS虚拟机器人圆形轨迹rviz可视化示例: './humanoid/ros_topic_demo/index.md'
- ROS2 Humble 路径点发布器: './waypoint_publisher.md'
- 人形机器人全身逆运动学与可操作性: './humanoid/whole_body_ik.md'
+ - G1 CostNet 栅格规划对照: './humanoid/g1_costnet_grid_planner/index.md'
- 人形机器人双臂末端可达域分析: './humanoid/humanoid_reachability.md'
- 人形机器人跌倒预判与保护性倒地: './humanoid/humanoid_fall_protection.md'
- 人形机器人 3D 碰撞体边界检测可视化: './humanoid/humanoid_collision_boundary.md'
diff --git a/src/humanoid/g1_costnet_grid_planner/.gitattributes b/src/humanoid/g1_costnet_grid_planner/.gitattributes
new file mode 100644
index 0000000000..f13e51c0ec
--- /dev/null
+++ b/src/humanoid/g1_costnet_grid_planner/.gitattributes
@@ -0,0 +1,2 @@
+*.py text eol=lf
+*.sh text eol=lf
diff --git a/src/humanoid/g1_costnet_grid_planner/CMakeLists.txt b/src/humanoid/g1_costnet_grid_planner/CMakeLists.txt
new file mode 100644
index 0000000000..084cb87281
--- /dev/null
+++ b/src/humanoid/g1_costnet_grid_planner/CMakeLists.txt
@@ -0,0 +1,6 @@
+cmake_minimum_required(VERSION 3.0.2)
+project(g1_costnet_grid_planner)
+find_package(catkin REQUIRED)
+catkin_package()
+catkin_install_python(PROGRAMS scripts/task_runner.py DESTINATION ${CATKIN_PACKAGE_BIN_DESTINATION})
+install(DIRECTORY launch DESTINATION ${CATKIN_PACKAGE_SHARE_DESTINATION})
diff --git a/src/humanoid/g1_costnet_grid_planner/README.md b/src/humanoid/g1_costnet_grid_planner/README.md
new file mode 100644
index 0000000000..77510e61b3
--- /dev/null
+++ b/src/humanoid/g1_costnet_grid_planner/README.md
@@ -0,0 +1,44 @@
+# G1 CostNet 栅格规划对照
+
+在独立栅格示例中比较 CostNet 辅助 A* 和曼哈顿 A* 的路径代价与搜索扩展数。
+
+## 环境与运行
+
+使用 Python 3.11。在模块目录运行 `python -m pip install -r requirements.txt`。
+
+```bash
+python main.py --output results
+```
+
+示例在一个合成栅格比较经典和神经辅助搜索,输出路线图和 JSON。网络仅处理相同 f 值的排序。可用 `--mode train_planner` 重训,用 `--mode plan_stdio` 接收占据栅格 JSON。
+
+## ROS 1 Noetic 启动
+
+本模块的 ROS 接口使用 ROS 1 Noetic(catkin、rospy)。在 Noetic 源码空间运行,将路径替换为实际值。
+
+```bash
+source /opt/ros/noetic/setup.bash
+export ROS_PACKAGE_PATH="$PWD:${ROS_PACKAGE_PATH:-}"
+roslaunch g1_costnet_grid_planner main.launch sim_python:=/path/to/python sim_entry:=/path/to/main.py
+```
+
+## 可复核结果
+
+| 方法或指标 | 路径代价(格) | 扩展节点 |
+| --- | ---: | ---: |
+| Manhattan A* | 48 | 401 |
+| 随仓库提供的 CostNet 辅助 A* | 48 | 49 |
+| 40 张合成测试图的 CostNet MAE | 0.756 | — |
+
+扩展数仅对应演示地图。推理按栅格各轴尺寸归一化相对坐标,各轴尺度至少为 32,以保持 32×32 训练输入及已有权重兼容。大图的搜索扩展数会随布局变化。
+
+```bash
+python main.py --mode train_planner --epochs 5 --samples 100 --seed 91 --weights results/trained
+python main.py --mode plan_stdio --robot-radius 0.3 --stride 3 < request.json
+```
+
+请求 JSON 包含 `map`(行优先的一维占据数组)、`width`、`height`、`resolution`(米/格)、`origin`、`start`、`goal`(二维世界坐标)。可用 `robot_radius`、`stride` 字段覆盖命令行默认值。重复在同一 Python 进程调用时复用权重,文件更新后重新加载。
+
+## 大模型使用声明
+
+使用 OpenAI Codex 辅助拆分、编程、验证和文档整理,提交者复核并对提交内容负责。
diff --git a/src/humanoid/g1_costnet_grid_planner/launch/main.launch b/src/humanoid/g1_costnet_grid_planner/launch/main.launch
new file mode 100644
index 0000000000..9adf983528
--- /dev/null
+++ b/src/humanoid/g1_costnet_grid_planner/launch/main.launch
@@ -0,0 +1,10 @@
+
+
+
+
+
+
+
+
+
+
diff --git a/src/humanoid/g1_costnet_grid_planner/main.py b/src/humanoid/g1_costnet_grid_planner/main.py
new file mode 100644
index 0000000000..981f457f0a
--- /dev/null
+++ b/src/humanoid/g1_costnet_grid_planner/main.py
@@ -0,0 +1,35 @@
+#!/usr/bin/env python3
+"""在独立栅格示例中比较 CostNet 辅助 A* 和曼哈顿 A* 的路径代价与搜索扩展数。"""
+import argparse
+import json
+from pathlib import Path
+
+
+def main():
+ parser = argparse.ArgumentParser(description=__doc__)
+ parser.add_argument('--mode', choices=['demo', 'train_planner', 'plan_stdio'], default='demo')
+ parser.add_argument('--output', type=Path)
+ parser.add_argument('--weights', type=Path, default=Path(__file__).resolve().parent / 'models')
+ parser.add_argument('--samples', type=int, default=240)
+ parser.add_argument('--epochs', type=int, default=20)
+ parser.add_argument('--seed', type=int, default=91)
+ parser.add_argument('--robot-radius', type=float, default=0.3)
+ parser.add_argument('--stride', type=int, default=3)
+ args = parser.parse_args()
+ if args.mode == 'train_planner' and (args.samples < 1 or args.epochs < 1 or args.seed < 0):
+ parser.error('training requires positive samples and epochs, and a nonnegative seed')
+ if args.mode == 'train_planner':
+ from patrol.planning import train_planner
+ print(json.dumps(train_planner(args.weights, epochs=args.epochs, sample_count=args.samples, seed=args.seed),indent=2))
+ return
+ if args.mode == 'plan_stdio':
+ import sys
+ from patrol.planning import plan_request
+ print(json.dumps(plan_request(json.load(sys.stdin),args.weights, robot_radius=args.robot_radius, stride=args.stride)))
+ return
+ from patrol.demo import run
+ print(json.dumps(run(args.weights, args.output or Path("results")), indent=2))
+
+
+if __name__ == '__main__':
+ main()
diff --git a/src/humanoid/g1_costnet_grid_planner/models/planner.pt b/src/humanoid/g1_costnet_grid_planner/models/planner.pt
new file mode 100644
index 0000000000..97c8a1cad3
Binary files /dev/null and b/src/humanoid/g1_costnet_grid_planner/models/planner.pt differ
diff --git a/src/humanoid/g1_costnet_grid_planner/package.xml b/src/humanoid/g1_costnet_grid_planner/package.xml
new file mode 100644
index 0000000000..92c0df4c7b
--- /dev/null
+++ b/src/humanoid/g1_costnet_grid_planner/package.xml
@@ -0,0 +1,9 @@
+
+
+ g1_costnet_grid_planner0.1.0
+ MuJoCo G1 visual floor perception, image control and ROS indoor navigation.
+ usernamekkyy
+ BSD-3-Clause
+ catkin
+ rospy
+
diff --git a/src/humanoid/g1_costnet_grid_planner/patrol/__init__.py b/src/humanoid/g1_costnet_grid_planner/patrol/__init__.py
new file mode 100644
index 0000000000..bc4204b4fb
--- /dev/null
+++ b/src/humanoid/g1_costnet_grid_planner/patrol/__init__.py
@@ -0,0 +1 @@
+"""Visual patrol on a physics simulated G1 humanoid."""
diff --git a/src/humanoid/g1_costnet_grid_planner/patrol/demo.py b/src/humanoid/g1_costnet_grid_planner/patrol/demo.py
new file mode 100644
index 0000000000..c118766e4b
--- /dev/null
+++ b/src/humanoid/g1_costnet_grid_planner/patrol/demo.py
@@ -0,0 +1,43 @@
+import json
+
+import numpy as np
+from PIL import Image, ImageDraw
+
+from .planning import astar, predict_costs
+
+
+def run(weights, output):
+ grid = np.zeros((32, 32), dtype=bool)
+ grid[5:26, 10:13] = True
+ grid[8:29, 21:24] = True
+ start, goal = (27, 3), (4, 28)
+ costs = predict_costs(grid, goal, weights)
+ baseline, baseline_expanded = astar(grid, start, goal)
+ assisted, assisted_expanded = astar(grid, start, goal, costs)
+ if len(assisted) != len(baseline):
+ raise RuntimeError('Search methods disagree on optimal path cost')
+ output.mkdir(parents=True, exist_ok=True)
+ report = {
+ 'path_cost': len(assisted) - 1,
+ 'baseline_expanded': baseline_expanded,
+ 'assisted_expanded': assisted_expanded,
+ 'baseline_route': baseline,
+ 'assisted_route': assisted,
+ }
+ (output / 'planner.json').write_text(json.dumps(report, indent=2), encoding='utf-8')
+ image = Image.new('RGB', (640, 355), 'white')
+ draw = ImageDraw.Draw(image)
+ for panel, (route, label) in enumerate((
+ (baseline, f'Manhattan A* ({baseline_expanded} expanded)'),
+ (assisted, f'CostNet A* ({assisted_expanded} expanded)'),
+ )):
+ left = 16 + 320 * panel
+ draw.text((left, 9), label, fill='#13202a')
+ for y, x in np.argwhere(grid):
+ draw.rectangle((left+x*9, 30+y*9, left+x*9+8, 30+y*9+8), fill='#333b42')
+ draw.line([(left+x*9+4, 34+y*9) for y,x in route], fill='#087a61', width=3)
+ for y, x in (start, goal):
+ draw.ellipse((left+x*9, 30+y*9, left+x*9+8, 30+y*9+8), fill='#b04827')
+ draw.text((16, 330), f'Path cost: {report["path_cost"]} cells for both methods', fill='#13202a')
+ image.save(output / 'planner.png')
+ return report
diff --git a/src/humanoid/g1_costnet_grid_planner/patrol/planning.py b/src/humanoid/g1_costnet_grid_planner/patrol/planning.py
new file mode 100644
index 0000000000..c48189657a
--- /dev/null
+++ b/src/humanoid/g1_costnet_grid_planner/patrol/planning.py
@@ -0,0 +1,184 @@
+"""Learned cost-to-go tie breaker for admissible grid A*.
+
+Manhattan f is the primary priority, so learned estimates cannot invalidate
+optimality on the four-connected grid. The network resolves equal-f plateaus.
+This is neural-assisted search, not an end-to-end replacement of graph search.
+"""
+import heapq
+import json
+from pathlib import Path
+from functools import lru_cache
+import math
+
+import numpy as np
+import torch
+from torch import nn
+
+
+class CostNet(nn.Module):
+ def __init__(self):
+ super().__init__()
+ self.net=nn.Sequential(nn.Conv2d(3,16,3,padding=1),nn.ReLU(),
+ nn.Conv2d(16,24,3,padding=2,dilation=2),nn.ReLU(),
+ nn.Conv2d(24,24,3,padding=4,dilation=4),nn.ReLU(),
+ nn.Conv2d(24,1,1),nn.Softplus())
+
+ def forward(self,x):
+ return self.net(x)
+
+
+def features(grid,goal):
+ yy,xx=np.indices(grid.shape)
+ height, width = grid.shape
+ return np.stack([grid, (xx - goal[1]) / max(32, width),
+ (yy - goal[0]) / max(32, height)]).astype(np.float32)
+
+
+def distances(grid,goal):
+ result=np.full(grid.shape,np.inf,dtype=np.float32)
+ result[goal]=0
+ queue=[(0,goal)]
+ while queue:
+ cost,point=heapq.heappop(queue)
+ if cost!=result[point]: continue
+ for dy,dx in [(1,0),(-1,0),(0,1),(0,-1)]:
+ q=(point[0]+dy,point[1]+dx)
+ if 0<=q[0]50
+ radius=int(np.ceil(robot_radius/resolution))
+ inflated=blocked.copy()
+ for dy in range(-min(radius, len(blocked)-1), min(radius, len(blocked)-1)+1):
+ for dx in range(-min(radius, blocked.shape[1]-1), min(radius, blocked.shape[1]-1)+1):
+ if dx*dx+dy*dy>radius*radius: continue
+ srcy=slice(max(0,-dy),min(len(blocked),len(blocked)-dy))
+ srcx=slice(max(0,-dx),min(blocked.shape[1],blocked.shape[1]-dx))
+ dsty=slice(max(0,dy),min(len(blocked),len(blocked)+dy))
+ dstx=slice(max(0,dx),min(blocked.shape[1],blocked.shape[1]+dx))
+ inflated[dsty,dstx]|=blocked[srcy,srcx]
+ hh,ww=inflated.shape[0]//stride,inflated.shape[1]//stride
+ grid=inflated[:hh*stride,:ww*stride].reshape(hh,stride,ww,stride).any(axis=(1,3))
+ cell=resolution*stride;ox,oy=request['origin']
+ def index(xy):
+ p=(math.floor((xy[1]-oy)/cell), math.floor((xy[0]-ox)/cell))
+ if not(0<=p[0]=1.5:
+ keep.append(route[i])
+ keep.append(route[-1])
+ world=[[ox+(p[1]+.5)*cell,oy+(p[0]+.5)*cell] for p in keep[1:]]
+ world[-1]=request['goal']
+ return dict(waypoints=world,expanded=expanded,baseline_expanded=baseline_expanded,
+ path_cost=len(route)-1,baseline_path_cost=len(baseline)-1,
+ unknown_policy='provisionally traversable; live local obstacle avoidance required',
+ method='Manhattan A* with neural equal-f tie breaking')
diff --git a/src/humanoid/g1_costnet_grid_planner/requirements.txt b/src/humanoid/g1_costnet_grid_planner/requirements.txt
new file mode 100644
index 0000000000..09a1ba72d0
--- /dev/null
+++ b/src/humanoid/g1_costnet_grid_planner/requirements.txt
@@ -0,0 +1,3 @@
+numpy>=1.26,<3
+torch>=2.5
+Pillow>=10
diff --git a/src/humanoid/g1_costnet_grid_planner/scripts/task_runner.py b/src/humanoid/g1_costnet_grid_planner/scripts/task_runner.py
new file mode 100755
index 0000000000..a48aee8ebd
--- /dev/null
+++ b/src/humanoid/g1_costnet_grid_planner/scripts/task_runner.py
@@ -0,0 +1,27 @@
+#!/usr/bin/env python3
+"""Launch standalone simulation/learning tasks through a ROS launch file."""
+import subprocess
+import rospy
+
+
+def main():
+ rospy.init_node('humanoid_task_runner')
+ args = [
+ rospy.get_param('~sim_python'), rospy.get_param('~sim_entry'),
+ '--weights', rospy.get_param('~weights'),
+ '--output', rospy.get_param('~output'),
+ ]
+ child = subprocess.Popen(args)
+ try:
+ while child.poll() is None and not rospy.is_shutdown():
+ rospy.rostime.wallsleep(.1)
+ if child.poll() not in (None,0):
+ raise RuntimeError('Simulation task failed with exit code %s'%child.returncode)
+ finally:
+ if child.poll() is None:
+ child.terminate()
+ child.wait(timeout=10)
+
+
+if __name__ == '__main__':
+ main()