SP
SP controlling player 0 (blue).
Human-partner qualitative comparisons across five MARL methods and three Co-π-tree variants on five Overcooked-AI layouts.
Compact coordination with dense interaction and fast handoffs between the human partner and the evaluated policy.
SP controlling player 0 (blue).
PBT controlling player 0 (blue).
FCP controlling player 0 (blue).
MEP controlling player 0 (blue).
COLE controlling player 0 (blue).
Co-π-tree controlling player 0 (blue).
Co-π-tree-PI controlling player 0 (blue).
Co-π-tree w/o P controlling player 0 (blue).
SP controlling player 1 (green).
PBT controlling player 1 (green).
FCP controlling player 1 (green).
MEP controlling player 1 (green).
COLE controlling player 1 (green).
Co-π-tree controlling player 1 (green).
Co-π-tree-PI controlling player 1 (green).
Co-π-tree w/o P controlling player 1 (green).
Long-horizon movement in the ring highlights how clearly each policy communicates routing intent to a human teammate.
SP controlling player 0 (blue).
PBT controlling player 0 (blue).
FCP controlling player 0 (blue).
MEP controlling player 0 (blue).
COLE controlling player 0 (blue).
Co-π-tree controlling player 0 (blue).
Co-π-tree-PI controlling player 0 (blue).
Co-π-tree w/o P controlling player 0 (blue).
SP controlling player 1 (green).
PBT controlling player 1 (green).
FCP controlling player 1 (green).
MEP controlling player 1 (green).
COLE controlling player 1 (green).
Co-π-tree controlling player 1 (green).
Co-π-tree-PI controlling player 1 (green).
Co-π-tree w/o P controlling player 1 (green).
Counter bottlenecks make this layout especially useful for comparing whether the human partner can read and trust each method's intent.
SP controlling player 0 (blue).
PBT controlling player 0 (blue).
FCP controlling player 0 (blue).
MEP controlling player 0 (blue).
COLE controlling player 0 (blue).
Co-π-tree controlling player 0 (blue).
Co-π-tree-PI controlling player 0 (blue).
Co-π-tree w/o P controlling player 0 (blue).
SP controlling player 1 (green).
PBT controlling player 1 (green).
FCP controlling player 1 (green).
MEP controlling player 1 (green).
COLE controlling player 1 (green).
Co-π-tree controlling player 1 (green).
Co-π-tree-PI controlling player 1 (green).
Co-π-tree w/o P controlling player 1 (green).
Specialized zones create strong role asymmetries, making it easy to inspect how each policy coordinates with a human under structural imbalance.
SP controlling player 0 (blue).
PBT controlling player 0 (blue).
FCP controlling player 0 (blue).
MEP controlling player 0 (blue).
COLE controlling player 0 (blue).
Co-π-tree controlling player 0 (blue).
Co-π-tree-PI controlling player 0 (blue).
Co-π-tree w/o P controlling player 0 (blue).
SP controlling player 1 (green).
PBT controlling player 1 (green).
FCP controlling player 1 (green).
MEP controlling player 1 (green).
COLE controlling player 1 (green).
Co-π-tree controlling player 1 (green).
Co-π-tree-PI controlling player 1 (green).
Co-π-tree w/o P controlling player 1 (green).
Separated workspaces emphasize complementary timing and reveal whether a human can easily anticipate each method's next move.
SP controlling player 0 (blue).
PBT controlling player 0 (blue).
FCP controlling player 0 (blue).
MEP controlling player 0 (blue).
COLE controlling player 0 (blue).
Co-π-tree controlling player 0 (blue).
Co-π-tree-PI controlling player 0 (blue).
Co-π-tree w/o P controlling player 0 (blue).
SP controlling player 1 (green).
PBT controlling player 1 (green).
FCP controlling player 1 (green).
MEP controlling player 1 (green).
COLE controlling player 1 (green).
Co-π-tree controlling player 1 (green).
Co-π-tree-PI controlling player 1 (green).
Co-π-tree w/o P controlling player 1 (green).