让多个智能体学会合作

本次运行结果

Dependency cache ready: /Users/sanbu/Documents/Codex/2026-10-20/revie/work/chapter-lab-runs/studio-zh/workspaces/multiagent/.cache/online-experiments/hands-on-modern-rl-experiment05-multiagent-games-4efbca48fa8c.ready
代码仓库： /Users/sanbu/Documents/Codex/2026-10-20/revie/work/chapter-lab-runs/studio-zh/workspaces/multiagent
实验运行目录： /Users/sanbu/Documents/Codex/2026-10-20/revie/work/chapter-lab-runs/studio-zh/workspaces/multiagent/modelscope-space/hands-on-modern-rl-experiment05-multiagent-games

可选任务：
  simple-spread        Cooperative Navigation · MPE/simple_spread_v3
  simple-tag           Predator & Prey · MPE/simple_tag_v3
  pistonball           Pistonball · Butterfly/pistonball_v6
  cooperative-pong     Cooperative Pong · Butterfly/cooperative_pong_v6

所选任务： Cooperative Navigation
算法： Parameter-sharing PPO
轮次安排：2 轮 × 1,008 步 = 2,016 总步数
每轮都会评估并保存一个可单独选择的模型。

Device: CPU
启动与在线 Studio 相同的训练流程…

objc[28483]: Class SDLApplication is implemented in both /Users/sanbu/Documents/Codex/2026-10-20/revie/work/studio-venv/lib/python3.12/site-packages/cv2/.dylibs/libSDL2-2.0.0.dylib (0x11d230890) and /Users/sanbu/Documents/Codex/2026-10-20/revie/work/studio-venv/lib/python3.12/site-packages/pygame/.dylibs/libSDL2-2.0.0.dylib (0x11ea9d2c8). This may cause spurious casting failures and mysterious crashes. One of the duplicates must be removed or renamed.
objc[28483]: Class SDLAppDelegate is implemented in both /Users/sanbu/Documents/Codex/2026-10-20/revie/work/studio-venv/lib/python3.12/site-packages/cv2/.dylibs/libSDL2-2.0.0.dylib (0x11d2308e0) and /Users/sanbu/Documents/Codex/2026-10-20/revie/work/studio-venv/lib/python3.12/site-packages/pygame/.dylibs/libSDL2-2.0.0.dylib (0x11ea9d318). This may cause spurious casting failures and mysterious crashes. One of the duplicates must be removed or renamed.
objc[28483]: Class SDLTranslatorResponder is implemented in both /Users/sanbu/Documents/Codex/2026-10-20/revie/work/studio-venv/lib/python3.12/site-packages/cv2/.dylibs/libSDL2-2.0.0.dylib (0x11d230958) and /Users/sanbu/Documents/Codex/2026-10-20/revie/work/studio-venv/lib/python3.12/site-packages/pygame/.dylibs/libSDL2-2.0.0.dylib (0x11ea9d390). This may cause spurious casting failures and mysterious crashes. One of the duplicates must be removed or renamed.
objc[28483]: Class SDLMessageBoxPresenter is implemented in both /Users/sanbu/Documents/Codex/2026-10-20/revie/work/studio-venv/lib/python3.12/site-packages/cv2/.dylibs/libSDL2-2.0.0.dylib (0x11d230980) and /Users/sanbu/Documents/Codex/2026-10-20/revie/work/studio-venv/lib/python3.12/site-packages/pygame/.dylibs/libSDL2-2.0.0.dylib (0x11ea9d3b8). This may cause spurious casting failures and mysterious crashes. One of the duplicates must be removed or renamed.
objc[28483]: Class SDL_cocoametalview is implemented in both /Users/sanbu/Documents/Codex/2026-10-20/revie/work/studio-venv/lib/python3.12/site-packages/cv2/.dylibs/libSDL2-2.0.0.dylib (0x11d2309d0) and /Users/sanbu/Documents/Codex/2026-10-20/revie/work/studio-venv/lib/python3.12/site-packages/pygame/.dylibs/libSDL2-2.0.0.dylib (0x11ea9d408). This may cause spurious casting failures and mysterious crashes. One of the duplicates must be removed or renamed.
objc[28483]: Class SDLOpenGLContext is implemented in both /Users/sanbu/Documents/Codex/2026-10-20/revie/work/studio-venv/lib/python3.12/site-packages/cv2/.dylibs/libSDL2-2.0.0.dylib (0x11d230a20) and /Users/sanbu/Documents/Codex/2026-10-20/revie/work/studio-venv/lib/python3.12/site-packages/pygame/.dylibs/libSDL2-2.0.0.dylib (0x11ea9d458). This may cause spurious casting failures and mysterious crashes. One of the duplicates must be removed or renamed.
objc[28483]: Class SDL_ShapeData is implemented in both /Users/sanbu/Documents/Codex/2026-10-20/revie/work/studio-venv/lib/python3.12/site-packages/cv2/.dylibs/libSDL2-2.0.0.dylib (0x11d230a98) and /Users/sanbu/Documents/Codex/2026-10-20/revie/work/studio-venv/lib/python3.12/site-packages/pygame/.dylibs/libSDL2-2.0.0.dylib (0x11ea9d4d0). This may cause spurious casting failures and mysterious crashes. One of the duplicates must be removed or renamed.
objc[28483]: Class SDL_CocoaClosure is implemented in both /Users/sanbu/Documents/Codex/2026-10-20/revie/work/studio-venv/lib/python3.12/site-packages/cv2/.dylibs/libSDL2-2.0.0.dylib (0x11d230ae8) and /Users/sanbu/Documents/Codex/2026-10-20/revie/work/studio-venv/lib/python3.12/site-packages/pygame/.dylibs/libSDL2-2.0.0.dylib (0x11ea9d520). This may cause spurious casting failures and mysterious crashes. One of the duplicates must be removed or renamed.
objc[28483]: Class SDL_VideoData is implemented in both /Users/sanbu/Documents/Codex/2026-10-20/revie/work/studio-venv/lib/python3.12/site-packages/cv2/.dylibs/libSDL2-2.0.0.dylib (0x11d230b38) and /Users/sanbu/Documents/Codex/2026-10-20/revie/work/studio-venv/lib/python3.12/site-packages/pygame/.dylibs/libSDL2-2.0.0.dylib (0x11ea9d570). This may cause spurious casting failures and mysterious crashes. One of the duplicates must be removed or renamed.
objc[28483]: Class SDL_WindowData is implemented in both /Users/sanbu/Documents/Codex/2026-10-20/revie/work/studio-venv/lib/python3.12/site-packages/cv2/.dylibs/libSDL2-2.0.0.dylib (0x11d230b88) and /Users/sanbu/Documents/Codex/2026-10-20/revie/work/studio-venv/lib/python3.12/site-packages/pygame/.dylibs/libSDL2-2.0.0.dylib (0x11ea9d5c0). This may cause spurious casting failures and mysterious crashes. One of the duplicates must be removed or renamed.
objc[28483]: Class SDLWindow is implemented in both /Users/sanbu/Documents/Codex/2026-10-20/revie/work/studio-venv/lib/python3.12/site-packages/cv2/.dylibs/libSDL2-2.0.0.dylib (0x11d230bb0) and /Users/sanbu/Documents/Codex/2026-10-20/revie/work/studio-venv/lib/python3.12/site-packages/pygame/.dylibs/libSDL2-2.0.0.dylib (0x11ea9d5e8). This may cause spurious casting failures and mysterious crashes. One of the duplicates must be removed or renamed.
objc[28483]: Class Cocoa_WindowListener is implemented in both /Users/sanbu/Documents/Codex/2026-10-20/revie/work/studio-venv/lib/python3.12/site-packages/cv2/.dylibs/libSDL2-2.0.0.dylib (0x11d230bd8) and /Users/sanbu/Documents/Codex/2026-10-20/revie/work/studio-venv/lib/python3.12/site-packages/pygame/.dylibs/libSDL2-2.0.0.dylib (0x11ea9d610). This may cause spurious casting failures and mysterious crashes. One of the duplicates must be removed or renamed.
objc[28483]: Class SDLView is implemented in both /Users/sanbu/Documents/Codex/2026-10-20/revie/work/studio-venv/lib/python3.12/site-packages/cv2/.dylibs/libSDL2-2.0.0.dylib (0x11d230c50) and /Users/sanbu/Documents/Codex/2026-10-20/revie/work/studio-venv/lib/python3.12/site-packages/pygame/.dylibs/libSDL2-2.0.0.dylib (0x11ea9d688). This may cause spurious casting failures and mysterious crashes. One of the duplicates must be removed or renamed.
objc[28483]: Class METAL_RenderData is implemented in both /Users/sanbu/Documents/Codex/2026-10-20/revie/work/studio-venv/lib/python3.12/site-packages/cv2/.dylibs/libSDL2-2.0.0.dylib (0x11d230cc8) and /Users/sanbu/Documents/Codex/2026-10-20/revie/work/studio-venv/lib/python3.12/site-packages/pygame/.dylibs/libSDL2-2.0.0.dylib (0x11ea9d700). This may cause spurious casting failures and mysterious crashes. One of the duplicates must be removed or renamed.
objc[28483]: Class METAL_TextureData is implemented in both /Users/sanbu/Documents/Codex/2026-10-20/revie/work/studio-venv/lib/python3.12/site-packages/cv2/.dylibs/libSDL2-2.0.0.dylib (0x11d230d18) and /Users/sanbu/Documents/Codex/2026-10-20/revie/work/studio-venv/lib/python3.12/site-packages/pygame/.dylibs/libSDL2-2.0.0.dylib (0x11ea9d750). This may cause spurious casting failures and mysterious crashes. One of the duplicates must be removed or renamed.
objc[28483]: Class SDL_RumbleMotor is implemented in both /Users/sanbu/Documents/Codex/2026-10-20/revie/work/studio-venv/lib/python3.12/site-packages/cv2/.dylibs/libSDL2-2.0.0.dylib (0x11d230d40) and /Users/sanbu/Documents/Codex/2026-10-20/revie/work/studio-venv/lib/python3.12/site-packages/pygame/.dylibs/libSDL2-2.0.0.dylib (0x11ea9d778). This may cause spurious casting failures and mysterious crashes. One of the duplicates must be removed or renamed.
objc[28483]: Class SDL_RumbleContext is implemented in both /Users/sanbu/Documents/Codex/2026-10-20/revie/work/studio-venv/lib/python3.12/site-packages/cv2/.dylibs/libSDL2-2.0.0.dylib (0x11d230d90) and /Users/sanbu/Documents/Codex/2026-10-20/revie/work/studio-venv/lib/python3.12/site-packages/pygame/.dylibs/libSDL2-2.0.0.dylib (0x11ea9d7c8). This may cause spurious casting failures and mysterious crashes. One of the duplicates must be removed or renamed.

Initialized shared PPO policy for 3 vector-agent slots

PPO update step=1,008 mean_agent_reward=-536.0193
SAVE epoch=1/2 model=policy.zip replay_frames=155

PPO update step=2,016 mean_agent_reward=-186.4740
SAVE epoch=2/2 model=policy.zip replay_frames=155

Saved 2 independently selectable multi-agent policies and replays

最终阶段： complete
最终得分： -186.4739675362404
已保存 2 个本次运行可选择的轮次模型。
模型索引： /Users/sanbu/Documents/Codex/2026-10-20/revie/work/chapter-lab-runs/studio-zh/workspaces/multiagent/modelscope-space/hands-on-modern-rl-experiment05-multiagent-games/artifacts/notebook-models.json

运行下面的推理单元格，选择轮次模型并查看该策略的回放。

VBox(children=(Dropdown(description='训练轮次模型：', layout=Layout(width='95%'), options=(('轮次 2/2 · 2,016 步 · 得分 -1…

默认选择最新轮次；也可以选择其他已保存模型，然后点击 回放所选策略.

{
  "score": -186.4739675362404,
  "steps": 2016,
  "epochs": 2,
  "checkpoints": 2,
  "preview": "learned-policy.gif"
}

{
  "score": -186.4739675362404,
  "steps": 2016,
  "epochs": 2,
  "checkpoints": 2,
  "preview": "learned-policy.gif"
}