122 lines
3.6 KiB
122 lines
3.6 KiB
import numpy as np
|
|
|
|
from gym_minigrid.minigrid import Goal, Grid, Lava, MiniGridEnv, MissionSpace
|
|
|
|
|
|
class LavaGapEnv(MiniGridEnv):
|
|
|
|
"""
|
|
### Description
|
|
|
|
The agent has to reach the green goal square at the opposite corner of the
|
|
room, and must pass through a narrow gap in a vertical strip of deadly lava.
|
|
Touching the lava terminate the episode with a zero reward. This environment
|
|
is useful for studying safety and safe exploration.
|
|
|
|
### Mission Space
|
|
|
|
Depending on the `obstacle_type` parameter:
|
|
- `Lava`: "avoid the lava and get to the green goal square"
|
|
- otherwise: "find the opening and get to the green goal square"
|
|
|
|
### Action Space
|
|
|
|
| Num | Name | Action |
|
|
|-----|--------------|--------------|
|
|
| 0 | left | Turn left |
|
|
| 1 | right | Turn right |
|
|
| 2 | forward | Move forward |
|
|
| 3 | pickup | Unused |
|
|
| 4 | drop | Unused |
|
|
| 5 | toggle | Unused |
|
|
| 6 | done | Unused |
|
|
|
|
### Observation Encoding
|
|
|
|
- Each tile is encoded as a 3 dimensional tuple:
|
|
`(OBJECT_IDX, COLOR_IDX, STATE)`
|
|
- `OBJECT_TO_IDX` and `COLOR_TO_IDX` mapping can be found in
|
|
[gym_minigrid/minigrid.py](gym_minigrid/minigrid.py)
|
|
- `STATE` refers to the door state with 0=open, 1=closed and 2=locked
|
|
|
|
### Rewards
|
|
|
|
A reward of '1' is given for success, and '0' for failure.
|
|
|
|
### Termination
|
|
|
|
The episode ends if any one of the following conditions is met:
|
|
|
|
1. The agent reaches the goal.
|
|
2. The agent falls into lava.
|
|
3. Timeout (see `max_steps`).
|
|
|
|
### Registered Configurations
|
|
|
|
S: size of map SxS.
|
|
|
|
- `MiniGrid-LavaGapS5-v0`
|
|
- `MiniGrid-LavaGapS6-v0`
|
|
- `MiniGrid-LavaGapS7-v0`
|
|
|
|
"""
|
|
|
|
def __init__(self, size, obstacle_type=Lava, **kwargs):
|
|
self.obstacle_type = obstacle_type
|
|
self.size = size
|
|
|
|
if obstacle_type == Lava:
|
|
mission_space = MissionSpace(
|
|
mission_func=lambda: "avoid the lava and get to the green goal square"
|
|
)
|
|
else:
|
|
mission_space = MissionSpace(
|
|
mission_func=lambda: "find the opening and get to the green goal square"
|
|
)
|
|
|
|
super().__init__(
|
|
mission_space=mission_space,
|
|
width=size,
|
|
height=size,
|
|
max_steps=4 * size * size,
|
|
# Set this to True for maximum speed
|
|
see_through_walls=False,
|
|
**kwargs
|
|
)
|
|
|
|
def _gen_grid(self, width, height):
|
|
assert width >= 5 and height >= 5
|
|
|
|
# Create an empty grid
|
|
self.grid = Grid(width, height)
|
|
|
|
# Generate the surrounding walls
|
|
self.grid.wall_rect(0, 0, width, height)
|
|
|
|
# Place the agent in the top-left corner
|
|
self.agent_pos = np.array((1, 1))
|
|
self.agent_dir = 0
|
|
|
|
# Place a goal square in the bottom-right corner
|
|
self.goal_pos = np.array((width - 2, height - 2))
|
|
self.put_obj(Goal(), *self.goal_pos)
|
|
|
|
# Generate and store random gap position
|
|
self.gap_pos = np.array(
|
|
(
|
|
self._rand_int(2, width - 2),
|
|
self._rand_int(1, height - 1),
|
|
)
|
|
)
|
|
|
|
# Place the obstacle wall
|
|
self.grid.vert_wall(self.gap_pos[0], 1, height - 2, self.obstacle_type)
|
|
|
|
# Put a hole in the wall
|
|
self.grid.set(*self.gap_pos, None)
|
|
|
|
self.mission = (
|
|
"avoid the lava and get to the green goal square"
|
|
if self.obstacle_type == Lava
|
|
else "find the opening and get to the green goal square"
|
|
)
|