You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
 
 
 
 
 
 

278 lines
9.0 KiB

#!/usr/bin/env python3
from __future__ import annotations
import re
from tqdm import tqdm
import gymnasium as gym
import numpy as np
import pygame
from gymnasium import Env
from minigrid.core.actions import Actions
from minigrid.core.state import to_state
from minigrid.minigrid_env import MiniGridEnv
from minigrid.wrappers import ImgObsWrapper, RGBImgPartialObsWrapper
def actionsToMiniGrid(actions):
mask = [0] * 7
for action in actions:
if "turn_left" in action:
mask[0] = 1
elif "turn_right" in action:
mask[1] = 1
elif "move" in action:
mask[2] = 1
elif "pickup" in action:
mask[3] = 1
elif "drop" in action:
mask[4] = 1
elif "toggle" in action:
mask[5] = 1
elif "done" in action:
mask[6] = 1
return mask
class Shield:
def __init__(self, shieldfile="current.shield"):
self.shieldfile = shieldfile
self.parse()
def parse(self):
self.shield = dict()
self.shield_raw = dict()
with open(self.shieldfile, "r") as shield:
shield = shield.readlines()
for line in tqdm(shield[3:-2]):
state_valuation = line[line.find("[")+1:line.find("]")]
actions = line[line.find("]")+1:]
ints = dict(re.findall(r'([a-zA-Z][_a-zA-Z0-9]+)=(-?[a-zA-Z0-9]+)', state_valuation))
booleans = re.findall(r'(\!?)([a-zA-Z][_a-zA-Z0-9]+)[\s\t]+', state_valuation)
booleans = {b[1]: False if b[0] == "!" else True for b in booleans}
actions = re.findall(r'{([a-zA-Z][_a-zA-Z0-9]+)}', actions)
if int(ints.get("clock", 0)) != 0:
continue
if int(ints.get("previousActionAgent", 3)) != 3:
continue
self.shield[to_state(ints, booleans)] = actionsToMiniGrid(actions)
self.shield_raw[to_state(ints, booleans)] = line[line.find("]")+1:]
def get_action_mask(self, state):
print(state)
try:
return self.shield[state]
except:
print("Unsafe State")
return [0.0] * 7
def get_action_mask_raw(self, state):
try:
return self.shield_raw[state]
except:
print("Not listed")
class ManualControl:
def __init__(
self,
env: Env,
seed=None,
random_agent=False,
shieldfile=None,
enforce=False
) -> None:
self.env = env
self.seed = seed
self.closed = False
self.random_agent = random_agent
if shieldfile is not None:
self.shield = Shield(shieldfile)
self.enforce= enforce
self.cumulative_reward = 0
def start(self):
"""Start the window display with blocking event loop"""
self.reset(self.seed)
while not self.closed:
if self.random_agent:
index = np.random.choice(7, 1)[0]
action = [Actions.left, Actions.right, Actions.forward, Actions.pickup, Actions.drop, Actions.toggle, Actions.done][index]
print(Actions(action), end=" ")
if hasattr(self, "shield") and self.enforce:
mask = self.shield.get_action_mask(self.env.get_symbolic_state())
if mask[Actions(action)] == 1.0:
self.step(action)
else:
print("blocked: ", Actions(action), end=" ")
else:
self.step(action)
print(" ")
else:
for event in pygame.event.get():
if event.type == pygame.QUIT:
self.env.close()
break
if event.type == pygame.KEYDOWN:
event.key = pygame.key.name(int(event.key))
self.key_handler(event)
def step(self, action: Actions):
_, reward, terminated, truncated, info = self.env.step(action)
self.cumulative_reward += reward
print(f"step={self.env.step_count}, reward={reward:.4f}, cumulative_reward={self.cumulative_reward:.4f}")
print(info)
if hasattr(self, "shield") and self.enforce:
symbolic_state = self.env.get_symbolic_state()
mask = self.shield.get_action_mask(symbolic_state)
print(mask)
print(self.shield.get_action_mask_raw(symbolic_state))
if terminated:
print("terminated!")
input("")
self.reset(self.seed)
elif truncated:
print("truncated!")
self.reset(self.seed)
else:
self.env.render()
def reset(self, seed=None):
self.env.reset(seed=seed)
if hasattr(self, "shield") and self.enforce:
symbolic_state = self.env.get_symbolic_state()
mask = self.shield.get_action_mask(symbolic_state)
print(mask)
print(self.shield.get_action_mask_raw(symbolic_state))
self.cumulative_reward = 0
self.env.render()
def key_handler(self, event):
key: str = event.key
print("pressed", key)
if key == "escape":
self.env.close()
return
if key == "backspace":
self.reset()
return
if key == "f12":
self.take_screenshot()
return
key_to_action = {
"left": Actions.left,
"right": Actions.right,
"up": Actions.forward,
"space": Actions.toggle,
"pageup": Actions.pickup,
"pagedown": Actions.drop,
"tab": Actions.pickup,
"left shift": Actions.drop,
"enter": Actions.done,
}
if key in key_to_action.keys():
action = key_to_action[key]
symbolic_state = self.env.get_symbolic_state()
if hasattr(self, "shield") and self.enforce:
mask = self.shield.get_action_mask(symbolic_state)
print(mask)
print(self.shield.get_action_mask_raw(symbolic_state))
if mask[Actions(action)] == 1.0:
self.step(action)
else:
print(key)
elif hasattr(self, "shield") and not self.enforce:
mask = self.shield.get_action_mask(symbolic_state)
print(mask)
print(self.shield.get_action_mask_raw(symbolic_state))
else:
self.step(action)
else:
print(key)
def take_screenshot(self):
import datetime
filename = f"{datetime.datetime.now().isoformat()}.png"
print(f"Saving a screenshot to '{filename}'")
window = self.env.window
screenshot = pygame.Surface(window.get_size())
screenshot.blit(window, (0,0))
pygame.image.save(screenshot, filename)
if __name__ == "__main__":
import argparse
parser = argparse.ArgumentParser()
parser.add_argument(
"--env-id",
type=str,
help="gym environment to load",
choices=gym.envs.registry.keys(),
default="MiniGrid-MultiRoom-N6-v0",
)
parser.add_argument(
"--seed",
type=int,
help="random seed to generate the environment with",
default=None,
)
parser.add_argument(
"--tile-size", type=int, help="size at which to render tiles", default=32
)
parser.add_argument(
"--agent-view",
action="store_true",
help="draw the agent sees (partially observable view)",
)
parser.add_argument(
"--agent-view-size",
type=int,
default=7,
help="set the number of grid spaces visible in agent-view ",
)
parser.add_argument(
"--screen-size",
type=int,
default="640",
help="set the resolution for pygame rendering (width and height)",
)
parser.add_argument(
"--random-agent",
action="store_true",
help="make the agent move around randomly"
)
parser.add_argument(
"--shield-file",
type=str,
help="shield file to parse and load",
)
parser.add_argument(
"--no-enforcement",
action="store_true",
help="do not enforce, but inform the user abouth shield violations"
)
args = parser.parse_args()
env: MiniGridEnv = gym.make(
args.env_id,
tile_size=args.tile_size,
render_mode="human",
agent_pov=args.agent_view,
agent_view_size=args.agent_view_size,
screen_size=args.screen_size,
)
if args.agent_view:
print("Using agent view")
env = RGBImgPartialObsWrapper(env, args.tile_size)
env = ImgObsWrapper(env)
#env.disable_random_start()
print(env.printGrid(init=True))
manual_control = ManualControl(env, seed=args.seed, random_agent=args.random_agent, shieldfile=args.shield_file, enforce=args.no_enforcement == False)
manual_control.start()