Reward Guide
-
Reach goal
+25
-
Move closer (Manhattan)
+1
-
Visit new cell
+0.25
-
Move farther (Manhattan)
−0.5
-
Invalid move (wall/obstacle)
−2
-
Timeout > 150 steps
−10
Observations are a normalized 4-value vector:
[agent_row, agent_col, goal_row, goal_col] scaled to [0, 1].
Action space is discrete with 4 actions:
0=UP
1=RIGHT
2=DOWN
3=LEFT.
The environment is Gymnasium-style with reset(),
step(action),
observation(), and info(), exposed as
window.env.