Description
The plate pickup shaped reward in the Overcooked environment (v1) is broken due to two bugs in jaxmarl/environments/overcooked/overcooked.py.
Bug 1: Incorrect dimension comparison (Line 617)
Current code:
plate_loc_layer = jnp.array(maze_map == OBJECT_TO_INDEX["plate"], dtype=jnp.uint8)
no_plates_on_counters = jnp.sum(plate_loc_layer) == 0
Problem:
- Compares against the entire 3D
maze_map instead of just the object channel
- Since
OBJECT_TO_INDEX["plate"] = 5 and COLOR_TO_INDEX["grey"] = 5, all walls [2, 5, 0] match on the color channel
- This causes
no_plates_on_counters to always be False, completely disabling the reward
Bug 2: Wrong inventory scope (Line 611)
Current code:
num_plates_in_inv = jnp.sum(inventory == OBJECT_TO_INDEX["plate"])
Problem:
- Only counts plates held by the current agent, not all agents
- Should check total plates across all agents like OvercookedV2 does
Description
The plate pickup shaped reward in the Overcooked environment (v1) is broken due to two bugs in
jaxmarl/environments/overcooked/overcooked.py.Bug 1: Incorrect dimension comparison (Line 617)
Current code:
plate_loc_layer = jnp.array(maze_map == OBJECT_TO_INDEX["plate"], dtype=jnp.uint8)
no_plates_on_counters = jnp.sum(plate_loc_layer) == 0
Problem:
maze_mapinstead of just the object channelOBJECT_TO_INDEX["plate"] = 5andCOLOR_TO_INDEX["grey"] = 5, all walls[2, 5, 0]match on the color channelno_plates_on_countersto always beFalse, completely disabling the rewardBug 2: Wrong inventory scope (Line 611)
Current code:
num_plates_in_inv = jnp.sum(inventory == OBJECT_TO_INDEX["plate"])
Problem: