Ad Hoc Teamwork: HFO
Ad hoc teamwork in the Half Field Offense (HFO) task in the 2D simulated RoboCup domain.
Original specification of the HFO task.
In this research, we investigate the performance of PLASTIC-Policy in the HFO domain. PLASTIC-Policy learns policies for cooperating with past teammates and selects from among these policies on the fly. In these videos, PLASTIC-Policy has learned about 7 previous types teammates. As a baseline, we compare the performance of the Combined Policy, where the ad hoc agent combines its observations about all 7 types of teammates to learn a new policy.
Contents
Limited Half Field Offense
In the limited half field offense task, 2 offensive agents play against 2 defensive agents, including the goalie. In these videos, the defense is using behaviors designed by Helios. We sample teammates coming from 2 different teams for the videos.
Cooperating with a teammate designed by Cyrus:
| Combined Policy: | PLASTIC-Policy: | |
| Off: 2 Def: 8 | Off: 6 Def: 4 |
Cooperating with a teammate from the agent2d code release:
| Combined Policy: | PLASTIC-Policy: | |
| Off: 2 Def: 8 | Off: 5 Def: 5 |
Full Half Field Offense
In the full half field offense task, 4 offensive agents play against 5 defensive agents, including the goalie. In these videos, the defense is using behaviors from the agent2d code release. We sample teammates coming from 2 different teams for the videos.
Cooperating with a teammate designed by Helios:
| Combined Policy: | PLASTIC-Policy: | |
| Off: 2 Def: 8 | Off: 5 Def: 5 |
Cooperating with a teammate designed by the Gliders:
| Combined Policy: | PLASTIC-Policy: | |
| Off: 3 Def: 7 | Off: 6 Def: 4 |