[Bug Report] Rainy Taxi is inconsistent in left and right #1510
Description
Describe the bug
The rainy taxi has an 80% change to move as intended, a 10% chance to go left of course, 10% chance to go right. The left/right should be relative to the direction of motion. From the original paper, "probability 0.1 it instead moves to the right (of the intended direction) and with probability 0.1 it moves to the left".
They are not consistent in the code, which reverses them for some directions. More importantly, the test for valid movements seems to assume movements are to the left and right based on a north heading, regardless of actual motion. This blocks some moves that should be allowed.
Example, in the image below with the taxi at (1,3), all movements are valid. However the left and right are not consistent
The transitions are:
If moving SOUTH, outcomes are
With probability 0.8 move ahead to (2, 3)
With probability 0.1 move left to (1, 2) <--- actually right
With probability 0.1 move right to (1, 4) <--- actually left
If moving NORTH, outcomes are
With probability 0.8 move ahead to (0, 3)
With probability 0.1 move left to (1, 2)
With probability 0.1 move right to (1, 4)
If moving EAST, outcomes are
With probability 0.8 move ahead to (1, 4)
With probability 0.1 move left to (2, 3) <--- actually right
With probability 0.1 move right to (0, 3) <--- actually left
If moving WEST, outcomes are
With probability 0.8 move ahead to (1, 2)
With probability 0.1 move left to (2, 3)
With probability 0.1 move right to (0, 3)
If the taxi is one step to the east (location 1,4) , the invalid moves should be: EAST (and left/right of east), NORTH right, SOUTH left. All other moves should be valid
The actual transitions are:
If moving SOUTH, outcomes are
With probability 0.8 move ahead to (2, 4)
With probability 0.1 move left to (1, 3)
With probability 0.1 move right to (1, 4) (no movement - move blocked) <-- wrong: left/right swapped
If moving NORTH, outcomes are
With probability 0.8 move ahead to (0, 4)
With probability 0.1 move left to (1, 3)
With probability 0.1 move right to (1, 4) (no movement - move blocked) <-- correct
If moving EAST, outcomes are
With probability 0.8 move ahead to (1, 4) (no movement - move blocked) <-- correct
With probability 0.1 move left to (1, 4) (no movement - move blocked) <-- correct
With probability 0.1 move right to (1, 4) (no movement - move blocked) <-- correct
If moving WEST, outcomes are
With probability 0.8 move ahead to (1, 3)
With probability 0.1 move left to (1, 4) (no movement - move blocked) <-- wrong: completely wrong
With probability 0.1 move right to (0, 4)
Code example
from gymnasium.envs.toy_text.taxi import TaxiEnv
env = TaxiEnv(is_rainy=True)
# taxi_loc = (1, 3) # example 1
taxi_loc = (1, 4) # example 2
pax_loc = 0
pax_dest = 1
state = env.encode(taxi_loc[0], taxi_loc[1], pax_loc, pax_dest)
actions = {'SOUTH': 0, 'NORTH': 1, 'EAST': 2, 'WEST': 3}
directions = ["ahead", "left ", "right"]
for action_name, action in actions.items():
print(f"If moving {action_name}, outcomes are")
for direction, (prob, new_state, _, _) in zip(directions, env.P[state][action]):
new_taxi_x, new_taxi_y, _, _ = env.decode(new_state)
new_taxi_loc = (new_taxi_x, new_taxi_y)
note = "(no movement - move blocked)" if new_taxi_loc == taxi_loc else ""
print(f" With probability {prob} move {direction} to {new_taxi_loc} {note}")System info
installed from latest source
gymnasium.__version__ = '1.2.3'
Additional context
I believe I know the cause of this and can fix it along with #1509
Does this require a version bump because the behavior changes?
Checklist
- I have checked that there is no similar issue in the repo