Windows正常运行的ML Agents项目在Linux执行启动失败求助

环境信息 #

• Python 3.8.10
• ml-agents: 0.27.0
• ml-agents-envs: 0.27.0
• Communicator API: 1.5.0
• PyTorch: 1.8.1+cu102
• Unity版本：2022.3.2f1（曾尝试2020.2.6f1，无改善）
• 目标系统：Ubuntu 20.04.6 LTS（SSH远程运行）

问题场景 #

Windows本地开发运行ML Agents仿真项目完全正常，切换到Ubuntu服务器提升仿真效率，将项目打包为Linux可执行文件后，通过SSH远程执行训练命令或直接用Python调用UnityEnvironment均失败，报错返回码1。

操作流程 #

1. Unity打包步骤：File->Build Settings->Windows, Mac, Linux->Linux，勾选Development Build后执行Build
2. 将打包后的文件上传至Ubuntu服务器
3. 执行训练命令：

mlagents-learn config/ppoagent.yaml --env=visibility_game_linux_build2022.x86_64 --no-graphics --force

4. 或尝试Python代码直接启动环境：

from mlagents_envs.environment import UnityEnvironment
env = UnityEnvironment(file_name='visibility_game_linux_build2022.x86_64')

报错信息 #

训练命令报错 #

[INFO] Learning was interrupted. Please wait while the graph is generated.
Traceback (most recent call last):
File "/home/rmarr/Documents/GflowsForSimulation_env/GflowsForSimulation_venv_real/bin/mlagents-learn", line 8, in
sys.exit(main())
File "/home/rmarr/Documents/GflowsForSimulation_env/GflowsForSimulation_venv_real/lib/python3.8/site-packages/mlagents/trainers/learn.py", line 250, in main
run_cli(parse_command_line())
File "/home/rmarr/Documents/GflowsForSimulation_env/GflowsForSimulation_venv_real/lib/python3.8/site-packages/mlagents/trainers/learn.py", line 246, in run_cli
run_training(run_seed, options)
File "/home/rmarr/Documents/GflowsForSimulation_env/GflowsForSimulation_venv_real/lib/python3.8/site-packages/mlagents/trainers/learn.py", line 125, in run_training
tc.start_learning(env_manager)
File "/home/rmarr/Documents/GflowsForSimulation_env/GflowsForSimulation_venv_real/lib/python3.8/site-packages/mlagents_envs/timers.py", line 305, in wrapped
return func(*args, **kwargs)
File "/home/rmarr/Documents/GflowsForSimulation_env/GflowsForSimulation_venv_real/lib/python3.8/site-packages/mlagents/trainers/trainer_controller.py", line 198, in start_learning
raise ex
File "/home/rmarr/Documents/GflowsForSimulation_env/GflowsForSimulation_venv_real/lib/python3.8/site-packages/mlagents/trainers/trainer_controller.py", line 173, in start_learning
self._reset_env(env_manager)
File "/home/rmarr/Documents/GflowsForSimulation_env/GflowsForSimulation_venv_real/lib/python3.8/site-packages/mlagents_envs/timers.py", line 305, in wrapped
return func(*args, **kwargs)
File "/home/rmarr/Documents/GflowsForSimulation_env/GflowsForSimulation_venv_real/lib/python3.8/site-packages/mlagents/trainers/trainer_controller.py", line 105, in _reset_env
env_manager.reset(config=new_config)
File "/home/rmarr/Documents/GflowsForSimulation_env/GflowsForSimulation_venv_real/lib/python3.8/site-packages/mlagents/trainers/env_manager.py", line 68, in reset
self.first_step_infos = self._reset_env(config)
File "/home/rmarr/Documents/GflowsForSimulation_env/GflowsForSimulation_venv_real/lib/python3.8/site-packages/mlagents/trainers/subprocess_env_manager.py", line 334, in _reset_env
ew.previous_step = EnvironmentStep(ew.recv().payload, ew.worker_id, {}, {})
File "/home/rmarr/Documents/GflowsForSimulation_env/GflowsForSimulation_venv_real/lib/python3.8/site-packages/mlagents/trainers/subprocess_env_manager.py", line 98, in recv
raise env_exception
mlagents_envs.exception.UnityEnvironmentException: Environment shut down with return code 1.

Python代码启动报错 #

Traceback (most recent call last):
File "", line 1, in
File "/home/rmarr/Documents/GflowsForSimulation_env/GflowsForSimulation_venv_real/lib/python3.8/site-packages/mlagents_envs/environment.py", line 223, in init
aca_output = self._send_academy_parameters(rl_init_parameters_in)
File "/home/rmarr/Documents/GflowsForSimulation_env/GflowsForSimulation_venv_real/lib/python3.8/site-packages/mlagents_envs/environment.py", line 477, in _send_academy_parameters
return self._communicator.initialize(inputs, self._poll_process)
File "/home/rmarr/Documents/GflowsForSimulation_env/GflowsForSimulation_venv_real/lib/python3.8/site-packages/mlagents_envs/rpc_communicator.py", line 121, in initialize
self.poll_for_timeout(poll_callback)
File "/home/rmarr/Documents/GflowsForSimulation_env/GflowsForSimulation_venv_real/lib/python3.8/site-packages/mlagents_envs/rpc_communicator.py", line 108, in poll_for_timeout
poll_callback()
File "/home/rmarr/Documents/GflowsForSimulation_env/GflowsForSimulation_venv_real/lib/python3.8/site-packages/mlagents_envs/environment.py", line 403, in _poll_process
raise UnityEnvironmentException(exc_msg)
mlagents_envs.exception.UnityEnvironmentException: Environment shut down with return code 1.

排查与解决步骤 #

1. 添加可执行权限
   确保Linux可执行文件有运行权限：

   chmod +x visibility_game_linux_build2022.x86_64
   

2. 检查依赖缺失
   直接运行可执行文件，查看具体依赖报错：

   ./visibility_game_linux_build2022.x86_64
   

   常见缺失依赖及安装命令：

   sudo apt-get install libgconf-2-4 libgtk-3-0 libx11-xcb1 libxcb-dri3-0 libnss3 libasound2
   

   无图形界面服务器需安装虚拟显示驱动：

   sudo apt-get install xvfb
   

   用xvfb启动训练：

   xvfb-run -a mlagents-learn config/ppoagent.yaml --env=visibility_game_linux_build2022.x86_64 --no-graphics --force
   

3. 修正Unity打包配置

   • 确认打包架构为Linux x86_64
   • 取消勾选Unity Remote Support
   • 移除项目中Windows特定代码（如System.Windows相关API），替换为跨平台实现

4. 版本兼容性校验
   尝试升级ml-agents到适配Unity 2022.x的0.28.x版本，或降级到对应稳定版本，确保Communicator API版本完全匹配

5. 设置显示环境变量
   在SSH会话中添加环境变量：

   export DISPLAY=:0.0
   

   或直接在启动命令前追加：

   DISPLAY=:0.0 mlagents-learn config/ppoagent.yaml --env=visibility_game_linux_build2022.x86_64 --no-graphics --force
   

6. 查看Unity崩溃日志
   日志路径通常为~/.config/unity3d/[你的公司名称]/[项目名称]/Player.log，通过日志定位具体崩溃原因（如资源缺失、脚本错误等）

内容的提问来源于stack exchange，提问作者Ryan Marr