Co-π-tree

Collaborating with AI

Layout-by-layout qualitative comparisons across SP, COLE, ProAgent, and Co-π-tree. For each method, one row shows player 0 (blue) control and the other shows player 1 (green) control.

4 methods 5 layouts player 0 / player 1 control
The videos visualize replay frames concatenated step by step for qualitative inspection and are not synchronized to wall-clock execution time. Measured runtime for the corresponding experiment is reported explicitly beneath each video.

Cramped Room

Dense contact and quick handoffs reveal whether collaboration policies remain legible after role swaps.

8 videos
SP
player 0 (blue)

Player 0 Control

SP controlling player 0 (blue).

Runtime: 30.9s
player 1 (green)

Player 1 Control

SP controlling player 1 (green).

Runtime: 24.1s
COLE
player 0 (blue)

Player 0 Control

COLE controlling player 0 (blue).

Runtime: 20.9s
player 1 (green)

Player 1 Control

COLE controlling player 1 (green).

Runtime: 20.9s
ProAgent
player 0 (blue)

Player 0 Control

ProAgent controlling player 0 (blue).

Runtime: 3231.0s
player 1 (green)

Player 1 Control

ProAgent controlling player 1 (green).

Runtime: 3432.6s
Co-π-tree
player 0 (blue)

Player 0 Control

Co-π-tree controlling player 0 (blue).

Runtime: 20.1s
player 1 (green)

Player 1 Control

Co-π-tree controlling player 1 (green).

Runtime: 18.7s

Coordination Ring

Longer-horizon routing in the ring layout makes role-conditioned movement patterns easy to compare across methods.

8 videos
SP
player 0 (blue)

Player 0 Control

SP controlling player 0 (blue).

Runtime: 22.1s
player 1 (green)

Player 1 Control

SP controlling player 1 (green).

Runtime: 21.0s
COLE
player 0 (blue)

Player 0 Control

COLE controlling player 0 (blue).

Runtime: 56.4s
player 1 (green)

Player 1 Control

COLE controlling player 1 (green).

Runtime: 56.5s
ProAgent
player 0 (blue)

Player 0 Control

ProAgent controlling player 0 (blue).

Runtime: 1370.9s
player 1 (green)

Player 1 Control

ProAgent controlling player 1 (green).

Runtime: 889.1s
Co-π-tree
player 0 (blue)

Player 0 Control

Co-π-tree controlling player 0 (blue).

Runtime: 18.9s
player 1 (green)

Player 1 Control

Co-π-tree controlling player 1 (green).

Runtime: 54.9s

Counter Circuit

Counter bottlenecks emphasize whether the acting agent communicates intent clearly after the controlled role changes.

8 videos
SP
player 0 (blue)

Player 0 Control

SP controlling player 0 (blue).

Runtime: 31.6s
player 1 (green)

Player 1 Control

SP controlling player 1 (green).

Runtime: 28.5s
COLE
player 0 (blue)

Player 0 Control

COLE controlling player 0 (blue).

Runtime: 37.7s
player 1 (green)

Player 1 Control

COLE controlling player 1 (green).

Runtime: 37.3s
ProAgent
player 0 (blue)

Player 0 Control

ProAgent controlling player 0 (blue).

Runtime: 790.2s
player 1 (green)

Player 1 Control

ProAgent controlling player 1 (green).

Runtime: 668.1s
Co-π-tree
player 0 (blue)

Player 0 Control

Co-π-tree controlling player 0 (blue).

Runtime: 64.1s
player 1 (green)

Player 1 Control

Co-π-tree controlling player 1 (green).

Runtime: 62.3s

Asymmetric Advantages

Asymmetric regions expose whether each method preserves specialization when the controlled agent changes from p0 to p1.

8 videos
SP
player 0 (blue)

Player 0 Control

SP controlling player 0 (blue).

Runtime: 31.0s
player 1 (green)

Player 1 Control

SP controlling player 1 (green).

Runtime: 28.0s
COLE
player 0 (blue)

Player 0 Control

COLE controlling player 0 (blue).

Runtime: 25.3s
player 1 (green)

Player 1 Control

COLE controlling player 1 (green).

Runtime: 25.3s
ProAgent
player 0 (blue)

Player 0 Control

ProAgent controlling player 0 (blue).

Runtime: 2096.5s
player 1 (green)

Player 1 Control

ProAgent controlling player 1 (green).

Runtime: 1312.2s
Co-π-tree
player 0 (blue)

Player 0 Control

Co-π-tree controlling player 0 (blue).

Runtime: 24.1s
player 1 (green)

Player 1 Control

Co-π-tree controlling player 1 (green).

Runtime: 23.6s

Forced Coordination

Split workspaces make it easy to see how each method handles complementary timing when the acting role is switched.

8 videos
SP
player 0 (blue)

Player 0 Control

SP controlling player 0 (blue).

Runtime: 25.4s
player 1 (green)

Player 1 Control

SP controlling player 1 (green).

Runtime: 23.6s
COLE
player 0 (blue)

Player 0 Control

COLE controlling player 0 (blue).

Runtime: 23.6s
player 1 (green)

Player 1 Control

COLE controlling player 1 (green).

Runtime: 23.9s
ProAgent
player 0 (blue)

Player 0 Control

ProAgent controlling player 0 (blue).

Runtime: 1927.0s
player 1 (green)

Player 1 Control

ProAgent controlling player 1 (green).

Runtime: 4550.1s
Co-π-tree
player 0 (blue)

Player 0 Control

Co-π-tree controlling player 0 (blue).

Runtime: 22.5s
player 1 (green)

Player 1 Control

Co-π-tree controlling player 1 (green).

Runtime: 21.8s