Spaces:
Sleeping
Sleeping
Upload folder using huggingface_hub
Browse files- teeunit_env/__init__.py +32 -0
- teeunit_env/client.py +251 -0
- teeunit_env/openenv.yaml +6 -0
- teeunit_env/pyproject.toml +47 -0
- teeunit_env/server/Dockerfile +38 -0
- teeunit_env/server/__init__.py +13 -0
- teeunit_env/server/app.py +309 -0
- teeunit_env/server/tee_environment.py +682 -0
teeunit_env/__init__.py
ADDED
|
@@ -0,0 +1,32 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Copyright (c) 2024 TeeUnit Project
|
| 2 |
+
# SPDX-License-Identifier: MIT
|
| 3 |
+
|
| 4 |
+
"""
|
| 5 |
+
TeeUnit OpenEnv Environment
|
| 6 |
+
|
| 7 |
+
An OpenEnv-compatible multi-agent arena environment wrapping the real Teeworlds 0.7.5 game
|
| 8 |
+
for LLM-based reinforcement learning training.
|
| 9 |
+
|
| 10 |
+
Example:
|
| 11 |
+
>>> from teeunit_env import TeeEnv
|
| 12 |
+
>>>
|
| 13 |
+
>>> with TeeEnv(base_url="http://localhost:8000") as env:
|
| 14 |
+
... env.reset()
|
| 15 |
+
... tools = env.list_tools()
|
| 16 |
+
... print([t.name for t in tools])
|
| 17 |
+
... # ['move', 'jump', 'aim', 'shoot', 'hook', 'get_status']
|
| 18 |
+
...
|
| 19 |
+
... # Get current game state
|
| 20 |
+
... status = env.call_tool("get_status")
|
| 21 |
+
... print(status)
|
| 22 |
+
...
|
| 23 |
+
... # Take actions
|
| 24 |
+
... env.call_tool("move", direction="right")
|
| 25 |
+
... env.call_tool("aim", x=500, y=300)
|
| 26 |
+
... env.call_tool("shoot", weapon=1)
|
| 27 |
+
"""
|
| 28 |
+
|
| 29 |
+
from .client import TeeEnv, TeeAction
|
| 30 |
+
|
| 31 |
+
__all__ = ["TeeEnv", "TeeAction"]
|
| 32 |
+
__version__ = "0.1.0"
|
teeunit_env/client.py
ADDED
|
@@ -0,0 +1,251 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Copyright (c) 2024 TeeUnit Project
|
| 2 |
+
# SPDX-License-Identifier: MIT
|
| 3 |
+
|
| 4 |
+
"""
|
| 5 |
+
TeeUnit Environment Client.
|
| 6 |
+
|
| 7 |
+
This module provides the client for connecting to a TeeUnit Environment server.
|
| 8 |
+
TeeEnv extends MCPToolClient to provide tool-calling style interactions.
|
| 9 |
+
|
| 10 |
+
Example:
|
| 11 |
+
>>> with TeeEnv(base_url="http://localhost:8000") as env:
|
| 12 |
+
... env.reset()
|
| 13 |
+
...
|
| 14 |
+
... # Discover tools
|
| 15 |
+
... tools = env.list_tools()
|
| 16 |
+
... print([t.name for t in tools]) # ['move', 'jump', 'aim', 'shoot', 'hook', 'get_status']
|
| 17 |
+
...
|
| 18 |
+
... # Get game state
|
| 19 |
+
... status = env.call_tool("get_status")
|
| 20 |
+
... print(status)
|
| 21 |
+
...
|
| 22 |
+
... # Take actions
|
| 23 |
+
... result = env.call_tool("move", direction="right")
|
| 24 |
+
... result = env.call_tool("shoot", weapon=2)
|
| 25 |
+
"""
|
| 26 |
+
|
| 27 |
+
from dataclasses import dataclass
|
| 28 |
+
from typing import Optional, Any, Dict
|
| 29 |
+
|
| 30 |
+
# Support both in-repo and standalone imports
|
| 31 |
+
try:
|
| 32 |
+
from openenv.core.mcp_client import MCPToolClient
|
| 33 |
+
except ImportError:
|
| 34 |
+
# Fallback for development/testing without openenv installed
|
| 35 |
+
class MCPToolClient:
|
| 36 |
+
"""Fallback MCPToolClient for development."""
|
| 37 |
+
|
| 38 |
+
def __init__(self, base_url: str = "http://localhost:8000", **kwargs):
|
| 39 |
+
self.base_url = base_url
|
| 40 |
+
self._connected = False
|
| 41 |
+
|
| 42 |
+
def __enter__(self):
|
| 43 |
+
self._connected = True
|
| 44 |
+
return self
|
| 45 |
+
|
| 46 |
+
def __exit__(self, *args):
|
| 47 |
+
self._connected = False
|
| 48 |
+
|
| 49 |
+
def reset(self, **kwargs):
|
| 50 |
+
pass
|
| 51 |
+
|
| 52 |
+
def list_tools(self):
|
| 53 |
+
return []
|
| 54 |
+
|
| 55 |
+
def call_tool(self, name: str, **kwargs):
|
| 56 |
+
return None
|
| 57 |
+
|
| 58 |
+
def step(self, action):
|
| 59 |
+
return None
|
| 60 |
+
|
| 61 |
+
def close(self):
|
| 62 |
+
self._connected = False
|
| 63 |
+
|
| 64 |
+
@classmethod
|
| 65 |
+
def from_docker_image(cls, image: str, **kwargs):
|
| 66 |
+
return cls(**kwargs)
|
| 67 |
+
|
| 68 |
+
@classmethod
|
| 69 |
+
def from_env(cls, env_id: str, **kwargs):
|
| 70 |
+
return cls(**kwargs)
|
| 71 |
+
|
| 72 |
+
|
| 73 |
+
@dataclass
|
| 74 |
+
class TeeAction:
|
| 75 |
+
"""
|
| 76 |
+
Action for the TeeUnit environment.
|
| 77 |
+
|
| 78 |
+
This is a convenience class for building actions to send to the environment.
|
| 79 |
+
The actual actions are sent via MCP tools (move, jump, aim, shoot, hook).
|
| 80 |
+
|
| 81 |
+
Attributes:
|
| 82 |
+
tool_name: The MCP tool to call
|
| 83 |
+
arguments: Arguments for the tool
|
| 84 |
+
"""
|
| 85 |
+
tool_name: str
|
| 86 |
+
arguments: Dict[str, Any]
|
| 87 |
+
|
| 88 |
+
@classmethod
|
| 89 |
+
def move(cls, direction: str = "none") -> "TeeAction":
|
| 90 |
+
"""Create a move action."""
|
| 91 |
+
return cls(tool_name="move", arguments={"direction": direction})
|
| 92 |
+
|
| 93 |
+
@classmethod
|
| 94 |
+
def jump(cls) -> "TeeAction":
|
| 95 |
+
"""Create a jump action."""
|
| 96 |
+
return cls(tool_name="jump", arguments={})
|
| 97 |
+
|
| 98 |
+
@classmethod
|
| 99 |
+
def aim(cls, x: int, y: int) -> "TeeAction":
|
| 100 |
+
"""Create an aim action."""
|
| 101 |
+
return cls(tool_name="aim", arguments={"x": x, "y": y})
|
| 102 |
+
|
| 103 |
+
@classmethod
|
| 104 |
+
def shoot(cls, weapon: int = -1) -> "TeeAction":
|
| 105 |
+
"""Create a shoot action."""
|
| 106 |
+
return cls(tool_name="shoot", arguments={"weapon": weapon})
|
| 107 |
+
|
| 108 |
+
@classmethod
|
| 109 |
+
def hook(cls) -> "TeeAction":
|
| 110 |
+
"""Create a hook action."""
|
| 111 |
+
return cls(tool_name="hook", arguments={})
|
| 112 |
+
|
| 113 |
+
@classmethod
|
| 114 |
+
def get_status(cls) -> "TeeAction":
|
| 115 |
+
"""Create a get_status action."""
|
| 116 |
+
return cls(tool_name="get_status", arguments={})
|
| 117 |
+
|
| 118 |
+
|
| 119 |
+
class TeeEnv(MCPToolClient):
|
| 120 |
+
"""
|
| 121 |
+
Client for the TeeUnit Environment.
|
| 122 |
+
|
| 123 |
+
This client provides a simple interface for interacting with the TeeUnit
|
| 124 |
+
Environment via MCP tools. It inherits all functionality from MCPToolClient:
|
| 125 |
+
- `list_tools()`: Discover available tools
|
| 126 |
+
- `call_tool(name, **kwargs)`: Call a tool by name
|
| 127 |
+
- `reset(**kwargs)`: Reset the environment
|
| 128 |
+
- `step(action)`: Execute an action
|
| 129 |
+
|
| 130 |
+
Available MCP Tools:
|
| 131 |
+
- `move(direction)`: Move left, right, or none
|
| 132 |
+
- `jump()`: Make the tee jump
|
| 133 |
+
- `aim(x, y)`: Aim at coordinates
|
| 134 |
+
- `shoot(weapon)`: Fire weapon (0-5 or -1 for current)
|
| 135 |
+
- `hook()`: Use grappling hook
|
| 136 |
+
- `get_status()`: Get game state as text
|
| 137 |
+
|
| 138 |
+
Example:
|
| 139 |
+
>>> # Connect to a running server
|
| 140 |
+
>>> with TeeEnv(base_url="http://localhost:8000") as env:
|
| 141 |
+
... env.reset()
|
| 142 |
+
...
|
| 143 |
+
... # List available tools
|
| 144 |
+
... tools = env.list_tools()
|
| 145 |
+
... for tool in tools:
|
| 146 |
+
... print(f"{tool.name}: {tool.description}")
|
| 147 |
+
...
|
| 148 |
+
... # Get game state
|
| 149 |
+
... status = env.call_tool("get_status")
|
| 150 |
+
... print(status)
|
| 151 |
+
...
|
| 152 |
+
... # Take actions
|
| 153 |
+
... env.call_tool("move", direction="right")
|
| 154 |
+
... env.call_tool("aim", x=500, y=300)
|
| 155 |
+
... env.call_tool("shoot", weapon=2)
|
| 156 |
+
|
| 157 |
+
Example with HuggingFace Space:
|
| 158 |
+
>>> # Connect to HuggingFace Space
|
| 159 |
+
>>> env = TeeEnv.from_env("ziadbc/teeunit-env")
|
| 160 |
+
>>> try:
|
| 161 |
+
... env.reset()
|
| 162 |
+
... status = env.call_tool("get_status")
|
| 163 |
+
... print(status)
|
| 164 |
+
... finally:
|
| 165 |
+
... env.close()
|
| 166 |
+
|
| 167 |
+
Example with Docker:
|
| 168 |
+
>>> # Automatically start container and connect
|
| 169 |
+
>>> env = TeeEnv.from_docker_image("teeunit-env:latest")
|
| 170 |
+
>>> try:
|
| 171 |
+
... env.reset()
|
| 172 |
+
... env.call_tool("move", direction="right")
|
| 173 |
+
... finally:
|
| 174 |
+
... env.close()
|
| 175 |
+
"""
|
| 176 |
+
|
| 177 |
+
def execute_action(self, action: TeeAction) -> Any:
|
| 178 |
+
"""
|
| 179 |
+
Execute a TeeAction.
|
| 180 |
+
|
| 181 |
+
Args:
|
| 182 |
+
action: TeeAction to execute
|
| 183 |
+
|
| 184 |
+
Returns:
|
| 185 |
+
Result from the MCP tool call
|
| 186 |
+
"""
|
| 187 |
+
return self.call_tool(action.tool_name, **action.arguments)
|
| 188 |
+
|
| 189 |
+
def get_status(self) -> str:
|
| 190 |
+
"""
|
| 191 |
+
Get the current game state as text.
|
| 192 |
+
|
| 193 |
+
Returns:
|
| 194 |
+
Text description of game state
|
| 195 |
+
"""
|
| 196 |
+
return self.call_tool("get_status")
|
| 197 |
+
|
| 198 |
+
def move(self, direction: str = "none") -> str:
|
| 199 |
+
"""
|
| 200 |
+
Move the tee.
|
| 201 |
+
|
| 202 |
+
Args:
|
| 203 |
+
direction: "left", "right", or "none"
|
| 204 |
+
|
| 205 |
+
Returns:
|
| 206 |
+
Result message
|
| 207 |
+
"""
|
| 208 |
+
return self.call_tool("move", direction=direction)
|
| 209 |
+
|
| 210 |
+
def jump(self) -> str:
|
| 211 |
+
"""
|
| 212 |
+
Make the tee jump.
|
| 213 |
+
|
| 214 |
+
Returns:
|
| 215 |
+
Result message
|
| 216 |
+
"""
|
| 217 |
+
return self.call_tool("jump")
|
| 218 |
+
|
| 219 |
+
def aim(self, x: int, y: int) -> str:
|
| 220 |
+
"""
|
| 221 |
+
Aim at coordinates.
|
| 222 |
+
|
| 223 |
+
Args:
|
| 224 |
+
x: Target X coordinate
|
| 225 |
+
y: Target Y coordinate
|
| 226 |
+
|
| 227 |
+
Returns:
|
| 228 |
+
Result message
|
| 229 |
+
"""
|
| 230 |
+
return self.call_tool("aim", x=x, y=y)
|
| 231 |
+
|
| 232 |
+
def shoot(self, weapon: int = -1) -> str:
|
| 233 |
+
"""
|
| 234 |
+
Fire weapon.
|
| 235 |
+
|
| 236 |
+
Args:
|
| 237 |
+
weapon: Weapon ID (0-5) or -1 for current weapon
|
| 238 |
+
|
| 239 |
+
Returns:
|
| 240 |
+
Result message
|
| 241 |
+
"""
|
| 242 |
+
return self.call_tool("shoot", weapon=weapon)
|
| 243 |
+
|
| 244 |
+
def hook(self) -> str:
|
| 245 |
+
"""
|
| 246 |
+
Use grappling hook.
|
| 247 |
+
|
| 248 |
+
Returns:
|
| 249 |
+
Result message
|
| 250 |
+
"""
|
| 251 |
+
return self.call_tool("hook")
|
teeunit_env/openenv.yaml
ADDED
|
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
spec_version: 1
|
| 2 |
+
name: teeunit_env
|
| 3 |
+
type: space
|
| 4 |
+
runtime: fastapi
|
| 5 |
+
app: server.app:app
|
| 6 |
+
port: 8000
|
teeunit_env/pyproject.toml
ADDED
|
@@ -0,0 +1,47 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[project]
|
| 2 |
+
name = "teeunit-env"
|
| 3 |
+
version = "0.1.0"
|
| 4 |
+
description = "OpenEnv-compatible Teeworlds arena environment for LLM RL training"
|
| 5 |
+
readme = "README.md"
|
| 6 |
+
license = "MIT"
|
| 7 |
+
authors = [
|
| 8 |
+
{ name = "TeeUnit Team", email = "ziadbc@users.noreply.github.com" }
|
| 9 |
+
]
|
| 10 |
+
requires-python = ">=3.10"
|
| 11 |
+
keywords = ["openenv", "teeworlds", "reinforcement-learning", "llm", "mcp"]
|
| 12 |
+
classifiers = [
|
| 13 |
+
"Development Status :: 3 - Alpha",
|
| 14 |
+
"Intended Audience :: Science/Research",
|
| 15 |
+
"License :: OSI Approved :: MIT License",
|
| 16 |
+
"Programming Language :: Python :: 3",
|
| 17 |
+
"Programming Language :: Python :: 3.10",
|
| 18 |
+
"Programming Language :: Python :: 3.11",
|
| 19 |
+
"Programming Language :: Python :: 3.12",
|
| 20 |
+
"Topic :: Scientific/Engineering :: Artificial Intelligence",
|
| 21 |
+
]
|
| 22 |
+
|
| 23 |
+
dependencies = [
|
| 24 |
+
"openenv-core>=0.2.1",
|
| 25 |
+
"fastmcp>=0.1.0",
|
| 26 |
+
"fastapi>=0.100.0",
|
| 27 |
+
"uvicorn>=0.23.0",
|
| 28 |
+
"pydantic>=2.0.0",
|
| 29 |
+
"websockets>=11.0",
|
| 30 |
+
]
|
| 31 |
+
|
| 32 |
+
[project.optional-dependencies]
|
| 33 |
+
dev = [
|
| 34 |
+
"pytest>=7.0",
|
| 35 |
+
"pytest-asyncio>=0.21",
|
| 36 |
+
"httpx>=0.24",
|
| 37 |
+
]
|
| 38 |
+
|
| 39 |
+
[project.scripts]
|
| 40 |
+
teeunit-server = "teeunit_env.server.app:main"
|
| 41 |
+
|
| 42 |
+
[build-system]
|
| 43 |
+
requires = ["hatchling"]
|
| 44 |
+
build-backend = "hatchling.build"
|
| 45 |
+
|
| 46 |
+
[tool.hatch.build.targets.wheel]
|
| 47 |
+
packages = ["teeunit_env"]
|
teeunit_env/server/Dockerfile
ADDED
|
@@ -0,0 +1,38 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# TeeUnit OpenEnv Environment
|
| 2 |
+
# Compatible with HuggingFace Spaces
|
| 3 |
+
|
| 4 |
+
FROM python:3.11-slim
|
| 5 |
+
|
| 6 |
+
WORKDIR /app
|
| 7 |
+
|
| 8 |
+
# Install system dependencies
|
| 9 |
+
RUN apt-get update && apt-get install -y --no-install-recommends \
|
| 10 |
+
build-essential \
|
| 11 |
+
&& rm -rf /var/lib/apt/lists/*
|
| 12 |
+
|
| 13 |
+
# Copy environment files
|
| 14 |
+
COPY teeunit_env/ ./teeunit_env/
|
| 15 |
+
COPY teeunit_env/pyproject.toml ./
|
| 16 |
+
|
| 17 |
+
# Install Python dependencies
|
| 18 |
+
RUN pip install --no-cache-dir \
|
| 19 |
+
openenv-core>=0.2.1 \
|
| 20 |
+
fastmcp>=0.1.0 \
|
| 21 |
+
fastapi>=0.100.0 \
|
| 22 |
+
uvicorn>=0.23.0 \
|
| 23 |
+
pydantic>=2.0.0 \
|
| 24 |
+
websockets>=11.0
|
| 25 |
+
|
| 26 |
+
# Expose port
|
| 27 |
+
EXPOSE 8000
|
| 28 |
+
|
| 29 |
+
# Set environment variables
|
| 30 |
+
ENV PYTHONPATH=/app
|
| 31 |
+
ENV PYTHONUNBUFFERED=1
|
| 32 |
+
|
| 33 |
+
# Health check
|
| 34 |
+
HEALTHCHECK --interval=30s --timeout=10s --start-period=5s --retries=3 \
|
| 35 |
+
CMD python -c "import urllib.request; urllib.request.urlopen('http://localhost:8000/health')" || exit 1
|
| 36 |
+
|
| 37 |
+
# Run the server
|
| 38 |
+
CMD ["uvicorn", "teeunit_env.server.app:app", "--host", "0.0.0.0", "--port", "8000"]
|
teeunit_env/server/__init__.py
ADDED
|
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Copyright (c) 2024 TeeUnit Project
|
| 2 |
+
# SPDX-License-Identifier: MIT
|
| 3 |
+
|
| 4 |
+
"""
|
| 5 |
+
TeeUnit OpenEnv Server
|
| 6 |
+
|
| 7 |
+
This module provides the server-side implementation of the TeeUnit environment,
|
| 8 |
+
exposing Teeworlds game control through MCP tools.
|
| 9 |
+
"""
|
| 10 |
+
|
| 11 |
+
from .tee_environment import TeeEnvironment
|
| 12 |
+
|
| 13 |
+
__all__ = ["TeeEnvironment"]
|
teeunit_env/server/app.py
ADDED
|
@@ -0,0 +1,309 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Copyright (c) 2024 TeeUnit Project
|
| 2 |
+
# SPDX-License-Identifier: MIT
|
| 3 |
+
|
| 4 |
+
"""
|
| 5 |
+
FastAPI application for the TeeUnit Environment.
|
| 6 |
+
|
| 7 |
+
This module creates an HTTP server that exposes the TeeEnvironment
|
| 8 |
+
over HTTP and WebSocket endpoints, compatible with MCPToolClient.
|
| 9 |
+
|
| 10 |
+
Usage:
|
| 11 |
+
# Development (with auto-reload):
|
| 12 |
+
uvicorn teeunit_env.server.app:app --reload --host 0.0.0.0 --port 8000
|
| 13 |
+
|
| 14 |
+
# Production:
|
| 15 |
+
uvicorn teeunit_env.server.app:app --host 0.0.0.0 --port 8000 --workers 4
|
| 16 |
+
|
| 17 |
+
# Or run directly:
|
| 18 |
+
python -m teeunit_env.server.app
|
| 19 |
+
"""
|
| 20 |
+
|
| 21 |
+
# Support both in-repo and standalone imports
|
| 22 |
+
try:
|
| 23 |
+
from openenv.core.env_server.http_server import create_app
|
| 24 |
+
from openenv.core.env_server.mcp_types import CallToolAction, CallToolObservation
|
| 25 |
+
|
| 26 |
+
from .tee_environment import TeeEnvironment
|
| 27 |
+
|
| 28 |
+
# Create the app with web interface
|
| 29 |
+
# Pass the class (factory) instead of an instance for WebSocket session support
|
| 30 |
+
# Use MCP types for action/observation since this is a pure MCP environment
|
| 31 |
+
app = create_app(
|
| 32 |
+
TeeEnvironment, CallToolAction, CallToolObservation, env_name="teeunit_env"
|
| 33 |
+
)
|
| 34 |
+
|
| 35 |
+
except ImportError:
|
| 36 |
+
# Fallback: Create a simple FastAPI app for development/testing
|
| 37 |
+
from fastapi import FastAPI, WebSocket, WebSocketDisconnect
|
| 38 |
+
from fastapi.middleware.cors import CORSMiddleware
|
| 39 |
+
from pydantic import BaseModel
|
| 40 |
+
from typing import Optional, Dict, Any
|
| 41 |
+
import json
|
| 42 |
+
|
| 43 |
+
# Import our environment
|
| 44 |
+
try:
|
| 45 |
+
from .tee_environment import TeeEnvironment
|
| 46 |
+
except ImportError:
|
| 47 |
+
from tee_environment import TeeEnvironment
|
| 48 |
+
|
| 49 |
+
app = FastAPI(
|
| 50 |
+
title="TeeUnit OpenEnv",
|
| 51 |
+
description="OpenEnv-compatible Teeworlds arena environment for LLM RL training",
|
| 52 |
+
version="0.1.0",
|
| 53 |
+
)
|
| 54 |
+
|
| 55 |
+
# Add CORS middleware
|
| 56 |
+
app.add_middleware(
|
| 57 |
+
CORSMiddleware,
|
| 58 |
+
allow_origins=["*"],
|
| 59 |
+
allow_credentials=True,
|
| 60 |
+
allow_methods=["*"],
|
| 61 |
+
allow_headers=["*"],
|
| 62 |
+
)
|
| 63 |
+
|
| 64 |
+
# Store environment instances per session
|
| 65 |
+
_environments: Dict[str, TeeEnvironment] = {}
|
| 66 |
+
|
| 67 |
+
class ResetRequest(BaseModel):
|
| 68 |
+
seed: Optional[int] = None
|
| 69 |
+
episode_id: Optional[str] = None
|
| 70 |
+
|
| 71 |
+
class ResetResponse(BaseModel):
|
| 72 |
+
status: str
|
| 73 |
+
episode_id: str
|
| 74 |
+
message: str
|
| 75 |
+
|
| 76 |
+
class ToolCallRequest(BaseModel):
|
| 77 |
+
tool_name: str
|
| 78 |
+
arguments: Dict[str, Any] = {}
|
| 79 |
+
|
| 80 |
+
class ToolCallResponse(BaseModel):
|
| 81 |
+
result: Any
|
| 82 |
+
reward: float
|
| 83 |
+
done: bool
|
| 84 |
+
metadata: Dict[str, Any] = {}
|
| 85 |
+
|
| 86 |
+
class ToolInfo(BaseModel):
|
| 87 |
+
name: str
|
| 88 |
+
description: str
|
| 89 |
+
parameters: Dict[str, Any] = {}
|
| 90 |
+
|
| 91 |
+
@app.get("/")
|
| 92 |
+
async def root():
|
| 93 |
+
"""Root endpoint with environment info."""
|
| 94 |
+
return {
|
| 95 |
+
"name": "TeeUnit OpenEnv",
|
| 96 |
+
"version": "0.1.0",
|
| 97 |
+
"description": "OpenEnv-compatible Teeworlds arena for LLM training",
|
| 98 |
+
"endpoints": {
|
| 99 |
+
"reset": "POST /reset",
|
| 100 |
+
"tools": "GET /tools",
|
| 101 |
+
"call_tool": "POST /call_tool",
|
| 102 |
+
"websocket": "WS /ws",
|
| 103 |
+
}
|
| 104 |
+
}
|
| 105 |
+
|
| 106 |
+
@app.get("/health")
|
| 107 |
+
async def health():
|
| 108 |
+
"""Health check endpoint."""
|
| 109 |
+
return {"status": "healthy"}
|
| 110 |
+
|
| 111 |
+
@app.post("/reset", response_model=ResetResponse)
|
| 112 |
+
async def reset(request: ResetRequest):
|
| 113 |
+
"""Reset the environment for a new episode."""
|
| 114 |
+
env = TeeEnvironment()
|
| 115 |
+
obs = env.reset(seed=request.seed, episode_id=request.episode_id)
|
| 116 |
+
|
| 117 |
+
session_id = obs.metadata.get("episode_id", "default")
|
| 118 |
+
_environments[session_id] = env
|
| 119 |
+
|
| 120 |
+
return ResetResponse(
|
| 121 |
+
status="ready",
|
| 122 |
+
episode_id=session_id,
|
| 123 |
+
message=obs.metadata.get("message", "Environment ready"),
|
| 124 |
+
)
|
| 125 |
+
|
| 126 |
+
@app.get("/tools")
|
| 127 |
+
async def list_tools():
|
| 128 |
+
"""List available MCP tools."""
|
| 129 |
+
return {
|
| 130 |
+
"tools": [
|
| 131 |
+
{
|
| 132 |
+
"name": "move",
|
| 133 |
+
"description": "Move the tee left, right, or none",
|
| 134 |
+
"parameters": {"direction": {"type": "string", "enum": ["left", "right", "none"]}},
|
| 135 |
+
},
|
| 136 |
+
{
|
| 137 |
+
"name": "jump",
|
| 138 |
+
"description": "Make the tee jump",
|
| 139 |
+
"parameters": {},
|
| 140 |
+
},
|
| 141 |
+
{
|
| 142 |
+
"name": "aim",
|
| 143 |
+
"description": "Aim at target coordinates",
|
| 144 |
+
"parameters": {
|
| 145 |
+
"x": {"type": "integer", "description": "Target X coordinate"},
|
| 146 |
+
"y": {"type": "integer", "description": "Target Y coordinate"},
|
| 147 |
+
},
|
| 148 |
+
},
|
| 149 |
+
{
|
| 150 |
+
"name": "shoot",
|
| 151 |
+
"description": "Fire the specified weapon",
|
| 152 |
+
"parameters": {
|
| 153 |
+
"weapon": {"type": "integer", "description": "Weapon ID (0-5) or -1 for current", "default": -1},
|
| 154 |
+
},
|
| 155 |
+
},
|
| 156 |
+
{
|
| 157 |
+
"name": "hook",
|
| 158 |
+
"description": "Use the grappling hook",
|
| 159 |
+
"parameters": {},
|
| 160 |
+
},
|
| 161 |
+
{
|
| 162 |
+
"name": "get_status",
|
| 163 |
+
"description": "Get current game state as text",
|
| 164 |
+
"parameters": {},
|
| 165 |
+
},
|
| 166 |
+
]
|
| 167 |
+
}
|
| 168 |
+
|
| 169 |
+
@app.post("/call_tool", response_model=ToolCallResponse)
|
| 170 |
+
async def call_tool(request: ToolCallRequest, session_id: str = "default"):
|
| 171 |
+
"""Call an MCP tool."""
|
| 172 |
+
env = _environments.get(session_id)
|
| 173 |
+
if env is None:
|
| 174 |
+
env = TeeEnvironment()
|
| 175 |
+
env.reset()
|
| 176 |
+
_environments[session_id] = env
|
| 177 |
+
|
| 178 |
+
# Get the MCP server from environment
|
| 179 |
+
mcp = env._mcp
|
| 180 |
+
|
| 181 |
+
# Call the tool
|
| 182 |
+
tool_name = request.tool_name
|
| 183 |
+
arguments = request.arguments
|
| 184 |
+
|
| 185 |
+
try:
|
| 186 |
+
# Use FastMCP's async call_tool method
|
| 187 |
+
tool_result = await mcp.call_tool(tool_name, arguments)
|
| 188 |
+
|
| 189 |
+
# Extract text result from ToolResult
|
| 190 |
+
if tool_result and tool_result.content:
|
| 191 |
+
result = tool_result.content[0].text if hasattr(tool_result.content[0], 'text') else str(tool_result.content[0])
|
| 192 |
+
else:
|
| 193 |
+
result = str(tool_result)
|
| 194 |
+
|
| 195 |
+
# Simulate tick and get reward
|
| 196 |
+
env._simulate_tick()
|
| 197 |
+
reward = env._calculate_reward()
|
| 198 |
+
|
| 199 |
+
# Check done
|
| 200 |
+
done = env._state.step_count >= env._max_steps
|
| 201 |
+
player = env._agents.get(env._current_agent_id)
|
| 202 |
+
if player and not player.is_alive:
|
| 203 |
+
done = True
|
| 204 |
+
|
| 205 |
+
return ToolCallResponse(
|
| 206 |
+
result=result,
|
| 207 |
+
reward=reward,
|
| 208 |
+
done=done,
|
| 209 |
+
metadata={
|
| 210 |
+
"step": env._state.step_count,
|
| 211 |
+
"tick": env._tick,
|
| 212 |
+
},
|
| 213 |
+
)
|
| 214 |
+
except Exception as e:
|
| 215 |
+
return ToolCallResponse(
|
| 216 |
+
result=f"Error: {str(e)}",
|
| 217 |
+
reward=0.0,
|
| 218 |
+
done=False,
|
| 219 |
+
metadata={"error": str(e)},
|
| 220 |
+
)
|
| 221 |
+
|
| 222 |
+
@app.websocket("/ws")
|
| 223 |
+
async def websocket_endpoint(websocket: WebSocket):
|
| 224 |
+
"""WebSocket endpoint for real-time interaction."""
|
| 225 |
+
await websocket.accept()
|
| 226 |
+
|
| 227 |
+
env = TeeEnvironment()
|
| 228 |
+
env.reset()
|
| 229 |
+
|
| 230 |
+
try:
|
| 231 |
+
while True:
|
| 232 |
+
data = await websocket.receive_text()
|
| 233 |
+
message = json.loads(data)
|
| 234 |
+
|
| 235 |
+
action_type = message.get("type", "call_tool")
|
| 236 |
+
|
| 237 |
+
if action_type == "reset":
|
| 238 |
+
obs = env.reset(
|
| 239 |
+
seed=message.get("seed"),
|
| 240 |
+
episode_id=message.get("episode_id"),
|
| 241 |
+
)
|
| 242 |
+
await websocket.send_json({
|
| 243 |
+
"type": "reset",
|
| 244 |
+
"status": "ready",
|
| 245 |
+
"episode_id": obs.metadata.get("episode_id"),
|
| 246 |
+
"message": obs.metadata.get("message"),
|
| 247 |
+
})
|
| 248 |
+
|
| 249 |
+
elif action_type == "list_tools":
|
| 250 |
+
await websocket.send_json({
|
| 251 |
+
"type": "tools",
|
| 252 |
+
"tools": [
|
| 253 |
+
{"name": "move", "description": "Move left/right/none"},
|
| 254 |
+
{"name": "jump", "description": "Jump"},
|
| 255 |
+
{"name": "aim", "description": "Aim at x,y"},
|
| 256 |
+
{"name": "shoot", "description": "Fire weapon"},
|
| 257 |
+
{"name": "hook", "description": "Use hook"},
|
| 258 |
+
{"name": "get_status", "description": "Get game state"},
|
| 259 |
+
],
|
| 260 |
+
})
|
| 261 |
+
|
| 262 |
+
elif action_type == "call_tool":
|
| 263 |
+
tool_name = message.get("tool_name")
|
| 264 |
+
arguments = message.get("arguments", {})
|
| 265 |
+
|
| 266 |
+
# Call tool using FastMCP's async call_tool method
|
| 267 |
+
mcp = env._mcp
|
| 268 |
+
try:
|
| 269 |
+
tool_result = await mcp.call_tool(tool_name, arguments)
|
| 270 |
+
if tool_result and tool_result.content:
|
| 271 |
+
result = tool_result.content[0].text if hasattr(tool_result.content[0], 'text') else str(tool_result.content[0])
|
| 272 |
+
else:
|
| 273 |
+
result = str(tool_result)
|
| 274 |
+
except Exception as e:
|
| 275 |
+
result = f"Error: {str(e)}"
|
| 276 |
+
|
| 277 |
+
# Simulate and get reward
|
| 278 |
+
env._simulate_tick()
|
| 279 |
+
reward = env._calculate_reward()
|
| 280 |
+
done = env._state.step_count >= env._max_steps
|
| 281 |
+
|
| 282 |
+
await websocket.send_json({
|
| 283 |
+
"type": "tool_result",
|
| 284 |
+
"tool_name": tool_name,
|
| 285 |
+
"result": result,
|
| 286 |
+
"reward": reward,
|
| 287 |
+
"done": done,
|
| 288 |
+
"step": env._state.step_count,
|
| 289 |
+
"tick": env._tick,
|
| 290 |
+
})
|
| 291 |
+
|
| 292 |
+
except WebSocketDisconnect:
|
| 293 |
+
pass
|
| 294 |
+
|
| 295 |
+
|
| 296 |
+
def main():
|
| 297 |
+
"""
|
| 298 |
+
Entry point for direct execution.
|
| 299 |
+
|
| 300 |
+
Usage:
|
| 301 |
+
python -m teeunit_env.server.app
|
| 302 |
+
uvicorn teeunit_env.server.app:app --host 0.0.0.0 --port 8000
|
| 303 |
+
"""
|
| 304 |
+
import uvicorn
|
| 305 |
+
uvicorn.run(app, host="0.0.0.0", port=8000)
|
| 306 |
+
|
| 307 |
+
|
| 308 |
+
if __name__ == "__main__":
|
| 309 |
+
main()
|
teeunit_env/server/tee_environment.py
ADDED
|
@@ -0,0 +1,682 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Copyright (c) 2024 TeeUnit Project
|
| 2 |
+
# SPDX-License-Identifier: MIT
|
| 3 |
+
|
| 4 |
+
"""
|
| 5 |
+
TeeUnit Environment Implementation.
|
| 6 |
+
|
| 7 |
+
A MCP environment that wraps the Teeworlds game for LLM-based RL training.
|
| 8 |
+
All interactions happen through MCP tools that translate to game actions.
|
| 9 |
+
|
| 10 |
+
MCP Tools:
|
| 11 |
+
- `move(direction)`: Move the tee left, right, or none
|
| 12 |
+
- `jump()`: Make the tee jump
|
| 13 |
+
- `aim(x, y)`: Aim at target coordinates
|
| 14 |
+
- `shoot(weapon)`: Fire the specified weapon
|
| 15 |
+
- `hook()`: Use the grappling hook
|
| 16 |
+
- `get_status()`: Get current game state as text
|
| 17 |
+
|
| 18 |
+
Example:
|
| 19 |
+
>>> from openenv.core.env_server.mcp_types import ListToolsAction, CallToolAction
|
| 20 |
+
>>> env = TeeEnvironment()
|
| 21 |
+
>>> env.reset()
|
| 22 |
+
>>>
|
| 23 |
+
>>> # List available tools
|
| 24 |
+
>>> obs = env.step(ListToolsAction())
|
| 25 |
+
>>> print([t.name for t in obs.tools]) # ["move", "jump", "aim", "shoot", "hook", "get_status"]
|
| 26 |
+
>>>
|
| 27 |
+
>>> # Get game state
|
| 28 |
+
>>> obs = env.step(CallToolAction(tool_name="get_status", arguments={}))
|
| 29 |
+
>>> print(obs.result)
|
| 30 |
+
"""
|
| 31 |
+
|
| 32 |
+
from typing import Any, Optional, Dict, List
|
| 33 |
+
from uuid import uuid4
|
| 34 |
+
import random
|
| 35 |
+
import math
|
| 36 |
+
|
| 37 |
+
# Support both in-repo and standalone imports
|
| 38 |
+
try:
|
| 39 |
+
from openenv.core.env_server.mcp_environment import MCPEnvironment
|
| 40 |
+
from openenv.core.env_server.types import Action, Observation, State
|
| 41 |
+
except ImportError:
|
| 42 |
+
# Fallback for development/testing
|
| 43 |
+
from dataclasses import dataclass
|
| 44 |
+
|
| 45 |
+
@dataclass
|
| 46 |
+
class State:
|
| 47 |
+
episode_id: str = ""
|
| 48 |
+
step_count: int = 0
|
| 49 |
+
|
| 50 |
+
@dataclass
|
| 51 |
+
class Observation:
|
| 52 |
+
done: bool = False
|
| 53 |
+
reward: float = 0.0
|
| 54 |
+
metadata: dict = None
|
| 55 |
+
|
| 56 |
+
def __post_init__(self):
|
| 57 |
+
if self.metadata is None:
|
| 58 |
+
self.metadata = {}
|
| 59 |
+
|
| 60 |
+
class Action:
|
| 61 |
+
pass
|
| 62 |
+
|
| 63 |
+
class MCPEnvironment:
|
| 64 |
+
def __init__(self, mcp):
|
| 65 |
+
self._mcp = mcp
|
| 66 |
+
|
| 67 |
+
def step(self, action, **kwargs):
|
| 68 |
+
return Observation()
|
| 69 |
+
|
| 70 |
+
from fastmcp import FastMCP
|
| 71 |
+
|
| 72 |
+
|
| 73 |
+
# Weapon definitions
|
| 74 |
+
WEAPONS = {
|
| 75 |
+
0: {"name": "hammer", "ammo": -1, "damage": 3},
|
| 76 |
+
1: {"name": "pistol", "ammo": 10, "damage": 1},
|
| 77 |
+
2: {"name": "shotgun", "ammo": 10, "damage": 3},
|
| 78 |
+
3: {"name": "grenade", "ammo": 10, "damage": 6},
|
| 79 |
+
4: {"name": "laser", "ammo": 10, "damage": 5},
|
| 80 |
+
5: {"name": "ninja", "ammo": -1, "damage": 9},
|
| 81 |
+
}
|
| 82 |
+
|
| 83 |
+
|
| 84 |
+
class GameAgent:
|
| 85 |
+
"""Represents a player/bot in the game."""
|
| 86 |
+
|
| 87 |
+
def __init__(self, agent_id: int):
|
| 88 |
+
self.agent_id = agent_id
|
| 89 |
+
self.x = 400.0 + random.uniform(-200, 200)
|
| 90 |
+
self.y = 300.0 + random.uniform(-100, 100)
|
| 91 |
+
self.vel_x = 0.0
|
| 92 |
+
self.vel_y = 0.0
|
| 93 |
+
self.health = 10
|
| 94 |
+
self.armor = 0
|
| 95 |
+
self.weapon = 1 # pistol
|
| 96 |
+
self.ammo = {w: WEAPONS[w]["ammo"] for w in WEAPONS}
|
| 97 |
+
self.direction = 1 # 1 = right, -1 = left
|
| 98 |
+
self.is_alive = True
|
| 99 |
+
self.score = 0
|
| 100 |
+
self.aim_x = self.x + 100
|
| 101 |
+
self.aim_y = self.y
|
| 102 |
+
self.is_hooking = False
|
| 103 |
+
self.is_grounded = True
|
| 104 |
+
|
| 105 |
+
def respawn(self):
|
| 106 |
+
"""Respawn at random location."""
|
| 107 |
+
self.x = 400.0 + random.uniform(-200, 200)
|
| 108 |
+
self.y = 300.0 + random.uniform(-100, 100)
|
| 109 |
+
self.vel_x = 0.0
|
| 110 |
+
self.vel_y = 0.0
|
| 111 |
+
self.health = 10
|
| 112 |
+
self.armor = 0
|
| 113 |
+
self.weapon = 1
|
| 114 |
+
self.ammo = {w: WEAPONS[w]["ammo"] for w in WEAPONS}
|
| 115 |
+
self.is_alive = True
|
| 116 |
+
self.is_hooking = False
|
| 117 |
+
|
| 118 |
+
|
| 119 |
+
class TeeEnvironment(MCPEnvironment):
|
| 120 |
+
"""
|
| 121 |
+
OpenEnv-compatible Teeworlds environment with MCP tool interface.
|
| 122 |
+
|
| 123 |
+
This environment provides a text-based interface for LLM agents to play
|
| 124 |
+
Teeworlds. The LLM receives game state as natural language descriptions
|
| 125 |
+
and issues commands through MCP tools.
|
| 126 |
+
|
| 127 |
+
For hackathon demo, this uses a simplified game simulation. For production,
|
| 128 |
+
it can be connected to the real Teeworlds server via bot_manager.
|
| 129 |
+
|
| 130 |
+
Example:
|
| 131 |
+
>>> with TeeEnv(base_url="http://localhost:8000") as env:
|
| 132 |
+
... env.reset()
|
| 133 |
+
... status = env.call_tool("get_status")
|
| 134 |
+
... env.call_tool("move", direction="right")
|
| 135 |
+
... env.call_tool("shoot", weapon=2)
|
| 136 |
+
"""
|
| 137 |
+
|
| 138 |
+
def __init__(
|
| 139 |
+
self,
|
| 140 |
+
num_agents: int = 4,
|
| 141 |
+
max_steps: int = 1000,
|
| 142 |
+
use_real_server: bool = False,
|
| 143 |
+
server_host: str = "127.0.0.1",
|
| 144 |
+
server_port: int = 8303,
|
| 145 |
+
):
|
| 146 |
+
"""
|
| 147 |
+
Initialize the TeeUnit environment.
|
| 148 |
+
|
| 149 |
+
Args:
|
| 150 |
+
num_agents: Number of agents in the arena
|
| 151 |
+
max_steps: Maximum steps per episode
|
| 152 |
+
use_real_server: If True, connect to real Teeworlds server
|
| 153 |
+
server_host: Teeworlds server host
|
| 154 |
+
server_port: Teeworlds server port
|
| 155 |
+
"""
|
| 156 |
+
# Create MCP server and define tools inline
|
| 157 |
+
mcp = FastMCP("teeunit_env")
|
| 158 |
+
|
| 159 |
+
# Store config
|
| 160 |
+
self._num_agents = num_agents
|
| 161 |
+
self._max_steps = max_steps
|
| 162 |
+
self._use_real_server = use_real_server
|
| 163 |
+
self._server_host = server_host
|
| 164 |
+
self._server_port = server_port
|
| 165 |
+
|
| 166 |
+
# Game state
|
| 167 |
+
self._agents: Dict[int, GameAgent] = {}
|
| 168 |
+
self._tick = 0
|
| 169 |
+
self._kill_events: List[dict] = []
|
| 170 |
+
self._current_agent_id = 0 # LLM controls agent 0
|
| 171 |
+
|
| 172 |
+
# Episode state
|
| 173 |
+
self._state = State(episode_id=str(uuid4()), step_count=0)
|
| 174 |
+
|
| 175 |
+
# Real server connection (optional)
|
| 176 |
+
self._bot_manager = None
|
| 177 |
+
|
| 178 |
+
# Define MCP tools
|
| 179 |
+
@mcp.tool
|
| 180 |
+
def move(direction: str) -> str:
|
| 181 |
+
"""
|
| 182 |
+
Move the tee horizontally.
|
| 183 |
+
|
| 184 |
+
Args:
|
| 185 |
+
direction: "left", "right", or "none"
|
| 186 |
+
|
| 187 |
+
Returns:
|
| 188 |
+
Result message describing the action taken
|
| 189 |
+
"""
|
| 190 |
+
agent = self._agents.get(self._current_agent_id)
|
| 191 |
+
if not agent or not agent.is_alive:
|
| 192 |
+
return "Cannot move: agent is dead"
|
| 193 |
+
|
| 194 |
+
if direction == "left":
|
| 195 |
+
agent.direction = -1
|
| 196 |
+
agent.vel_x = max(agent.vel_x - 5, -15)
|
| 197 |
+
return f"Moving left. Velocity: ({agent.vel_x:.1f}, {agent.vel_y:.1f})"
|
| 198 |
+
elif direction == "right":
|
| 199 |
+
agent.direction = 1
|
| 200 |
+
agent.vel_x = min(agent.vel_x + 5, 15)
|
| 201 |
+
return f"Moving right. Velocity: ({agent.vel_x:.1f}, {agent.vel_y:.1f})"
|
| 202 |
+
else:
|
| 203 |
+
agent.vel_x *= 0.8 # friction
|
| 204 |
+
return f"Stopped. Velocity: ({agent.vel_x:.1f}, {agent.vel_y:.1f})"
|
| 205 |
+
|
| 206 |
+
@mcp.tool
|
| 207 |
+
def jump() -> str:
|
| 208 |
+
"""
|
| 209 |
+
Make the tee jump. Can double-jump in the air.
|
| 210 |
+
|
| 211 |
+
Returns:
|
| 212 |
+
Result message describing the jump
|
| 213 |
+
"""
|
| 214 |
+
agent = self._agents.get(self._current_agent_id)
|
| 215 |
+
if not agent or not agent.is_alive:
|
| 216 |
+
return "Cannot jump: agent is dead"
|
| 217 |
+
|
| 218 |
+
if agent.is_grounded:
|
| 219 |
+
agent.vel_y = -12
|
| 220 |
+
agent.is_grounded = False
|
| 221 |
+
return f"Jumped! Velocity: ({agent.vel_x:.1f}, {agent.vel_y:.1f})"
|
| 222 |
+
else:
|
| 223 |
+
# Air jump (weaker)
|
| 224 |
+
agent.vel_y = -8
|
| 225 |
+
return f"Air jumped! Velocity: ({agent.vel_x:.1f}, {agent.vel_y:.1f})"
|
| 226 |
+
|
| 227 |
+
@mcp.tool
|
| 228 |
+
def aim(x: int, y: int) -> str:
|
| 229 |
+
"""
|
| 230 |
+
Aim at target coordinates.
|
| 231 |
+
|
| 232 |
+
Args:
|
| 233 |
+
x: Target X coordinate
|
| 234 |
+
y: Target Y coordinate
|
| 235 |
+
|
| 236 |
+
Returns:
|
| 237 |
+
Result message confirming aim direction
|
| 238 |
+
"""
|
| 239 |
+
agent = self._agents.get(self._current_agent_id)
|
| 240 |
+
if not agent or not agent.is_alive:
|
| 241 |
+
return "Cannot aim: agent is dead"
|
| 242 |
+
|
| 243 |
+
agent.aim_x = x
|
| 244 |
+
agent.aim_y = y
|
| 245 |
+
|
| 246 |
+
# Calculate angle for display
|
| 247 |
+
dx = x - agent.x
|
| 248 |
+
dy = y - agent.y
|
| 249 |
+
angle = math.atan2(dy, dx) * 180 / math.pi
|
| 250 |
+
distance = math.sqrt(dx*dx + dy*dy)
|
| 251 |
+
|
| 252 |
+
return f"Aiming at ({x}, {y}). Angle: {angle:.1f} deg, Distance: {distance:.1f} units"
|
| 253 |
+
|
| 254 |
+
@mcp.tool
|
| 255 |
+
def shoot(weapon: int = -1) -> str:
|
| 256 |
+
"""
|
| 257 |
+
Fire the current or specified weapon.
|
| 258 |
+
|
| 259 |
+
Args:
|
| 260 |
+
weapon: Weapon ID (0=hammer, 1=pistol, 2=shotgun, 3=grenade, 4=laser, 5=ninja).
|
| 261 |
+
Use -1 for current weapon.
|
| 262 |
+
|
| 263 |
+
Returns:
|
| 264 |
+
Result message describing the shot and any hits
|
| 265 |
+
"""
|
| 266 |
+
agent = self._agents.get(self._current_agent_id)
|
| 267 |
+
if not agent or not agent.is_alive:
|
| 268 |
+
return "Cannot shoot: agent is dead"
|
| 269 |
+
|
| 270 |
+
# Switch weapon if specified
|
| 271 |
+
if weapon >= 0 and weapon <= 5:
|
| 272 |
+
agent.weapon = weapon
|
| 273 |
+
|
| 274 |
+
wpn = WEAPONS[agent.weapon]
|
| 275 |
+
wpn_name = wpn["name"]
|
| 276 |
+
|
| 277 |
+
# Check ammo
|
| 278 |
+
if wpn["ammo"] > 0 and agent.ammo[agent.weapon] <= 0:
|
| 279 |
+
return f"Out of ammo for {wpn_name}!"
|
| 280 |
+
|
| 281 |
+
# Use ammo
|
| 282 |
+
if wpn["ammo"] > 0:
|
| 283 |
+
agent.ammo[agent.weapon] -= 1
|
| 284 |
+
|
| 285 |
+
# Check for hits on other agents
|
| 286 |
+
hits = []
|
| 287 |
+
for other_id, other in self._agents.items():
|
| 288 |
+
if other_id == self._current_agent_id or not other.is_alive:
|
| 289 |
+
continue
|
| 290 |
+
|
| 291 |
+
# Simple hit detection based on aim
|
| 292 |
+
dx = other.x - agent.x
|
| 293 |
+
dy = other.y - agent.y
|
| 294 |
+
distance = math.sqrt(dx*dx + dy*dy)
|
| 295 |
+
|
| 296 |
+
aim_dx = agent.aim_x - agent.x
|
| 297 |
+
aim_dy = agent.aim_y - agent.y
|
| 298 |
+
aim_dist = math.sqrt(aim_dx*aim_dx + aim_dy*aim_dy)
|
| 299 |
+
|
| 300 |
+
if aim_dist > 0:
|
| 301 |
+
# Check if enemy is roughly in line of fire
|
| 302 |
+
dot = (dx * aim_dx + dy * aim_dy) / (aim_dist * max(distance, 1))
|
| 303 |
+
|
| 304 |
+
# Hit probability based on weapon and distance
|
| 305 |
+
hit_range = 400 if agent.weapon != 0 else 50 # hammer short range
|
| 306 |
+
if distance < hit_range and dot > 0.8:
|
| 307 |
+
# Hit!
|
| 308 |
+
damage = wpn["damage"]
|
| 309 |
+
other.health -= damage
|
| 310 |
+
other.armor = max(0, other.armor - damage // 2)
|
| 311 |
+
|
| 312 |
+
if other.health <= 0:
|
| 313 |
+
other.is_alive = False
|
| 314 |
+
agent.score += 1
|
| 315 |
+
self._kill_events.append({
|
| 316 |
+
"killer_id": self._current_agent_id,
|
| 317 |
+
"victim_id": other_id,
|
| 318 |
+
"weapon": agent.weapon,
|
| 319 |
+
"tick": self._tick,
|
| 320 |
+
})
|
| 321 |
+
hits.append(f"KILLED Player {other_id} with {wpn_name}!")
|
| 322 |
+
else:
|
| 323 |
+
hits.append(f"Hit Player {other_id} for {damage} damage ({other.health}HP remaining)")
|
| 324 |
+
|
| 325 |
+
ammo_str = f"({agent.ammo[agent.weapon]} ammo)" if wpn["ammo"] > 0 else ""
|
| 326 |
+
if hits:
|
| 327 |
+
return f"Fired {wpn_name} {ammo_str}. " + " ".join(hits)
|
| 328 |
+
else:
|
| 329 |
+
return f"Fired {wpn_name} {ammo_str}. No hits."
|
| 330 |
+
|
| 331 |
+
@mcp.tool
|
| 332 |
+
def hook() -> str:
|
| 333 |
+
"""
|
| 334 |
+
Use the grappling hook in the aim direction.
|
| 335 |
+
The hook can grab walls or enemies to pull yourself toward them.
|
| 336 |
+
|
| 337 |
+
Returns:
|
| 338 |
+
Result message describing hook action
|
| 339 |
+
"""
|
| 340 |
+
agent = self._agents.get(self._current_agent_id)
|
| 341 |
+
if not agent or not agent.is_alive:
|
| 342 |
+
return "Cannot hook: agent is dead"
|
| 343 |
+
|
| 344 |
+
agent.is_hooking = not agent.is_hooking
|
| 345 |
+
|
| 346 |
+
if agent.is_hooking:
|
| 347 |
+
# Pull toward aim point
|
| 348 |
+
dx = agent.aim_x - agent.x
|
| 349 |
+
dy = agent.aim_y - agent.y
|
| 350 |
+
dist = math.sqrt(dx*dx + dy*dy)
|
| 351 |
+
if dist > 0:
|
| 352 |
+
agent.vel_x += (dx / dist) * 3
|
| 353 |
+
agent.vel_y += (dy / dist) * 3
|
| 354 |
+
return f"Hook deployed! Pulling toward ({agent.aim_x}, {agent.aim_y})"
|
| 355 |
+
else:
|
| 356 |
+
return "Hook released."
|
| 357 |
+
|
| 358 |
+
@mcp.tool
|
| 359 |
+
def get_status() -> str:
|
| 360 |
+
"""
|
| 361 |
+
Get the current game state as a text description.
|
| 362 |
+
|
| 363 |
+
Returns:
|
| 364 |
+
Detailed text description of current game state including:
|
| 365 |
+
- Your position, health, weapon, ammo
|
| 366 |
+
- Visible enemies with positions and health
|
| 367 |
+
- Recent events (kills, deaths)
|
| 368 |
+
"""
|
| 369 |
+
agent = self._agents.get(self._current_agent_id)
|
| 370 |
+
|
| 371 |
+
lines = []
|
| 372 |
+
lines.append(f"=== Teeworlds Game State (Tick {self._tick}) ===")
|
| 373 |
+
lines.append("")
|
| 374 |
+
|
| 375 |
+
if not agent or not agent.is_alive:
|
| 376 |
+
lines.append("STATUS: DEAD - Waiting for respawn...")
|
| 377 |
+
lines.append("")
|
| 378 |
+
else:
|
| 379 |
+
lines.append(f"Position: ({agent.x:.0f}, {agent.y:.0f}) | Velocity: ({agent.vel_x:.1f}, {agent.vel_y:.1f})")
|
| 380 |
+
lines.append(f"Health: {agent.health}/10 | Armor: {agent.armor}/10")
|
| 381 |
+
|
| 382 |
+
wpn = WEAPONS[agent.weapon]
|
| 383 |
+
ammo_str = str(agent.ammo[agent.weapon]) if wpn["ammo"] > 0 else "infinite"
|
| 384 |
+
lines.append(f"Weapon: {wpn['name']} ({ammo_str} ammo)")
|
| 385 |
+
lines.append(f"Score: {agent.score} kills")
|
| 386 |
+
lines.append(f"Aim: ({agent.aim_x:.0f}, {agent.aim_y:.0f})")
|
| 387 |
+
lines.append("")
|
| 388 |
+
|
| 389 |
+
# Other players
|
| 390 |
+
enemies = []
|
| 391 |
+
for other_id, other in self._agents.items():
|
| 392 |
+
if other_id == self._current_agent_id:
|
| 393 |
+
continue
|
| 394 |
+
|
| 395 |
+
if other.is_alive:
|
| 396 |
+
dx = other.x - agent.x if agent else other.x
|
| 397 |
+
dy = other.y - agent.y if agent else other.y
|
| 398 |
+
dist = math.sqrt(dx*dx + dy*dy)
|
| 399 |
+
wpn_name = WEAPONS[other.weapon]["name"]
|
| 400 |
+
enemies.append(
|
| 401 |
+
f" - Player {other_id}: pos({other.x:.0f}, {other.y:.0f}), "
|
| 402 |
+
f"{other.health}HP, {wpn_name}, {dist:.0f} units away"
|
| 403 |
+
)
|
| 404 |
+
else:
|
| 405 |
+
enemies.append(f" - Player {other_id}: DEAD")
|
| 406 |
+
|
| 407 |
+
if enemies:
|
| 408 |
+
lines.append("OTHER PLAYERS:")
|
| 409 |
+
lines.extend(enemies)
|
| 410 |
+
else:
|
| 411 |
+
lines.append("OTHER PLAYERS: None")
|
| 412 |
+
lines.append("")
|
| 413 |
+
|
| 414 |
+
# Recent kills
|
| 415 |
+
recent = self._kill_events[-5:] if self._kill_events else []
|
| 416 |
+
if recent:
|
| 417 |
+
lines.append("RECENT EVENTS:")
|
| 418 |
+
for event in recent:
|
| 419 |
+
killer = event["killer_id"]
|
| 420 |
+
victim = event["victim_id"]
|
| 421 |
+
wpn_name = WEAPONS[event["weapon"]]["name"]
|
| 422 |
+
if killer == self._current_agent_id:
|
| 423 |
+
lines.append(f" - You killed Player {victim} with {wpn_name}")
|
| 424 |
+
elif victim == self._current_agent_id:
|
| 425 |
+
lines.append(f" - Player {killer} killed you with {wpn_name}")
|
| 426 |
+
else:
|
| 427 |
+
lines.append(f" - Player {killer} killed Player {victim} with {wpn_name}")
|
| 428 |
+
lines.append("")
|
| 429 |
+
|
| 430 |
+
lines.append("AVAILABLE ACTIONS: move, jump, aim, shoot, hook, get_status")
|
| 431 |
+
|
| 432 |
+
return "\n".join(lines)
|
| 433 |
+
|
| 434 |
+
# Store MCP reference and pass to base class
|
| 435 |
+
self._mcp = mcp
|
| 436 |
+
super().__init__(mcp)
|
| 437 |
+
|
| 438 |
+
def reset(
|
| 439 |
+
self,
|
| 440 |
+
seed: Optional[int] = None,
|
| 441 |
+
episode_id: Optional[str] = None,
|
| 442 |
+
**kwargs: Any,
|
| 443 |
+
) -> Observation:
|
| 444 |
+
"""
|
| 445 |
+
Reset the environment for a new episode.
|
| 446 |
+
|
| 447 |
+
Args:
|
| 448 |
+
seed: Optional random seed
|
| 449 |
+
episode_id: Optional episode ID
|
| 450 |
+
**kwargs: Additional reset options
|
| 451 |
+
|
| 452 |
+
Returns:
|
| 453 |
+
Observation indicating the environment is ready
|
| 454 |
+
"""
|
| 455 |
+
if seed is not None:
|
| 456 |
+
random.seed(seed)
|
| 457 |
+
|
| 458 |
+
# Reset episode state
|
| 459 |
+
self._state = State(
|
| 460 |
+
episode_id=episode_id or str(uuid4()),
|
| 461 |
+
step_count=0,
|
| 462 |
+
)
|
| 463 |
+
self._tick = 0
|
| 464 |
+
self._kill_events = []
|
| 465 |
+
|
| 466 |
+
# Initialize agents
|
| 467 |
+
self._agents = {}
|
| 468 |
+
for i in range(self._num_agents):
|
| 469 |
+
self._agents[i] = GameAgent(i)
|
| 470 |
+
|
| 471 |
+
# Get initial status
|
| 472 |
+
status = self._get_status_text()
|
| 473 |
+
|
| 474 |
+
return Observation(
|
| 475 |
+
done=False,
|
| 476 |
+
reward=0.0,
|
| 477 |
+
metadata={
|
| 478 |
+
"status": "ready",
|
| 479 |
+
"message": status,
|
| 480 |
+
"episode_id": self._state.episode_id,
|
| 481 |
+
},
|
| 482 |
+
)
|
| 483 |
+
|
| 484 |
+
def _get_status_text(self) -> str:
|
| 485 |
+
"""Generate current game status text."""
|
| 486 |
+
agent = self._agents.get(self._current_agent_id)
|
| 487 |
+
|
| 488 |
+
lines = []
|
| 489 |
+
lines.append(f"=== Teeworlds Game State (Tick {self._tick}) ===")
|
| 490 |
+
|
| 491 |
+
if agent and agent.is_alive:
|
| 492 |
+
lines.append(f"Position: ({agent.x:.0f}, {agent.y:.0f})")
|
| 493 |
+
lines.append(f"Health: {agent.health}/10 | Armor: {agent.armor}/10")
|
| 494 |
+
wpn = WEAPONS[agent.weapon]
|
| 495 |
+
lines.append(f"Weapon: {wpn['name']}")
|
| 496 |
+
lines.append(f"Score: {agent.score} kills")
|
| 497 |
+
else:
|
| 498 |
+
lines.append("STATUS: DEAD")
|
| 499 |
+
|
| 500 |
+
return "\n".join(lines)
|
| 501 |
+
|
| 502 |
+
def _simulate_tick(self):
|
| 503 |
+
"""Simulate one game tick (physics, AI, etc.)."""
|
| 504 |
+
self._tick += 1
|
| 505 |
+
|
| 506 |
+
for agent in self._agents.values():
|
| 507 |
+
if not agent.is_alive:
|
| 508 |
+
continue
|
| 509 |
+
|
| 510 |
+
# Apply gravity
|
| 511 |
+
agent.vel_y += 0.5
|
| 512 |
+
|
| 513 |
+
# Apply velocity
|
| 514 |
+
agent.x += agent.vel_x
|
| 515 |
+
agent.y += agent.vel_y
|
| 516 |
+
|
| 517 |
+
# Ground collision (simple)
|
| 518 |
+
if agent.y > 500:
|
| 519 |
+
agent.y = 500
|
| 520 |
+
agent.vel_y = 0
|
| 521 |
+
agent.is_grounded = True
|
| 522 |
+
|
| 523 |
+
# Wall collision
|
| 524 |
+
agent.x = max(50, min(750, agent.x))
|
| 525 |
+
|
| 526 |
+
# Friction
|
| 527 |
+
agent.vel_x *= 0.95
|
| 528 |
+
|
| 529 |
+
# Simple AI for non-player agents
|
| 530 |
+
if agent.agent_id != self._current_agent_id:
|
| 531 |
+
self._simple_ai(agent)
|
| 532 |
+
|
| 533 |
+
def _simple_ai(self, agent: GameAgent):
|
| 534 |
+
"""Simple AI behavior for non-player agents."""
|
| 535 |
+
# Random movement
|
| 536 |
+
if random.random() < 0.1:
|
| 537 |
+
agent.vel_x += random.uniform(-3, 3)
|
| 538 |
+
|
| 539 |
+
# Random jump
|
| 540 |
+
if agent.is_grounded and random.random() < 0.05:
|
| 541 |
+
agent.vel_y = -10
|
| 542 |
+
agent.is_grounded = False
|
| 543 |
+
|
| 544 |
+
# Aim at player
|
| 545 |
+
player = self._agents.get(self._current_agent_id)
|
| 546 |
+
if player and player.is_alive:
|
| 547 |
+
agent.aim_x = player.x
|
| 548 |
+
agent.aim_y = player.y
|
| 549 |
+
|
| 550 |
+
# Occasionally shoot
|
| 551 |
+
if random.random() < 0.02:
|
| 552 |
+
dx = player.x - agent.x
|
| 553 |
+
dy = player.y - agent.y
|
| 554 |
+
dist = math.sqrt(dx*dx + dy*dy)
|
| 555 |
+
|
| 556 |
+
if dist < 300:
|
| 557 |
+
# Attack player
|
| 558 |
+
wpn = WEAPONS[agent.weapon]
|
| 559 |
+
if wpn["ammo"] < 0 or agent.ammo[agent.weapon] > 0:
|
| 560 |
+
if wpn["ammo"] > 0:
|
| 561 |
+
agent.ammo[agent.weapon] -= 1
|
| 562 |
+
|
| 563 |
+
# Check hit (simplified)
|
| 564 |
+
if dist < 200 and random.random() < 0.3:
|
| 565 |
+
damage = wpn["damage"]
|
| 566 |
+
player.health -= damage
|
| 567 |
+
|
| 568 |
+
if player.health <= 0:
|
| 569 |
+
player.is_alive = False
|
| 570 |
+
agent.score += 1
|
| 571 |
+
self._kill_events.append({
|
| 572 |
+
"killer_id": agent.agent_id,
|
| 573 |
+
"victim_id": self._current_agent_id,
|
| 574 |
+
"weapon": agent.weapon,
|
| 575 |
+
"tick": self._tick,
|
| 576 |
+
})
|
| 577 |
+
|
| 578 |
+
def _step_impl(
|
| 579 |
+
self,
|
| 580 |
+
action: Action,
|
| 581 |
+
timeout_s: Optional[float] = None,
|
| 582 |
+
**kwargs: Any,
|
| 583 |
+
) -> Observation:
|
| 584 |
+
"""
|
| 585 |
+
Handle non-MCP actions.
|
| 586 |
+
|
| 587 |
+
Args:
|
| 588 |
+
action: The action to execute
|
| 589 |
+
timeout_s: Optional timeout
|
| 590 |
+
**kwargs: Additional arguments
|
| 591 |
+
|
| 592 |
+
Returns:
|
| 593 |
+
Observation with error for unknown action types
|
| 594 |
+
"""
|
| 595 |
+
return Observation(
|
| 596 |
+
done=False,
|
| 597 |
+
reward=0.0,
|
| 598 |
+
metadata={
|
| 599 |
+
"error": f"Unknown action type: {type(action).__name__}. "
|
| 600 |
+
"Use ListToolsAction or CallToolAction for MCP interactions."
|
| 601 |
+
},
|
| 602 |
+
)
|
| 603 |
+
|
| 604 |
+
def step(
|
| 605 |
+
self,
|
| 606 |
+
action: Action,
|
| 607 |
+
timeout_s: Optional[float] = None,
|
| 608 |
+
**kwargs: Any,
|
| 609 |
+
) -> Observation:
|
| 610 |
+
"""
|
| 611 |
+
Execute a step in the environment.
|
| 612 |
+
|
| 613 |
+
Args:
|
| 614 |
+
action: The MCP action to execute
|
| 615 |
+
timeout_s: Optional timeout
|
| 616 |
+
**kwargs: Additional arguments
|
| 617 |
+
|
| 618 |
+
Returns:
|
| 619 |
+
Observation from the action execution
|
| 620 |
+
"""
|
| 621 |
+
# Increment step count
|
| 622 |
+
self._state.step_count += 1
|
| 623 |
+
|
| 624 |
+
# Simulate game tick
|
| 625 |
+
self._simulate_tick()
|
| 626 |
+
|
| 627 |
+
# Calculate reward
|
| 628 |
+
reward = self._calculate_reward()
|
| 629 |
+
|
| 630 |
+
# Check done
|
| 631 |
+
done = self._state.step_count >= self._max_steps
|
| 632 |
+
|
| 633 |
+
# Check if all enemies dead (win condition)
|
| 634 |
+
enemies_alive = sum(1 for a in self._agents.values()
|
| 635 |
+
if a.agent_id != self._current_agent_id and a.is_alive)
|
| 636 |
+
if enemies_alive == 0:
|
| 637 |
+
done = True
|
| 638 |
+
reward += 10.0 # Win bonus
|
| 639 |
+
|
| 640 |
+
# Check if player dead
|
| 641 |
+
player = self._agents.get(self._current_agent_id)
|
| 642 |
+
if player and not player.is_alive:
|
| 643 |
+
done = True
|
| 644 |
+
reward -= 5.0 # Death penalty
|
| 645 |
+
|
| 646 |
+
# Let the base class handle MCP actions
|
| 647 |
+
obs = super().step(action, timeout_s=timeout_s, **kwargs)
|
| 648 |
+
|
| 649 |
+
# Update observation with reward and done
|
| 650 |
+
obs.reward = reward
|
| 651 |
+
obs.done = done
|
| 652 |
+
obs.metadata["step"] = self._state.step_count
|
| 653 |
+
obs.metadata["tick"] = self._tick
|
| 654 |
+
|
| 655 |
+
return obs
|
| 656 |
+
|
| 657 |
+
def _calculate_reward(self) -> float:
|
| 658 |
+
"""Calculate reward for current step."""
|
| 659 |
+
reward = 0.0
|
| 660 |
+
|
| 661 |
+
player = self._agents.get(self._current_agent_id)
|
| 662 |
+
if not player:
|
| 663 |
+
return reward
|
| 664 |
+
|
| 665 |
+
# Survival bonus
|
| 666 |
+
if player.is_alive:
|
| 667 |
+
reward += 0.01
|
| 668 |
+
|
| 669 |
+
# Kill bonus (from recent events)
|
| 670 |
+
for event in self._kill_events:
|
| 671 |
+
if event["tick"] == self._tick:
|
| 672 |
+
if event["killer_id"] == self._current_agent_id:
|
| 673 |
+
reward += 1.0
|
| 674 |
+
elif event["victim_id"] == self._current_agent_id:
|
| 675 |
+
reward -= 0.5
|
| 676 |
+
|
| 677 |
+
return reward
|
| 678 |
+
|
| 679 |
+
@property
|
| 680 |
+
def state(self) -> State:
|
| 681 |
+
"""Get the current environment state."""
|
| 682 |
+
return self._state
|