Simulation results in GA3C-CADRL about cadrl_ros HOT 2 OPEN

There should be code in the gym repo to replicate these figures. In particular the small test suite script should contain the right scenarios, but you might need to set the number of agents since there are some different test scenarios for each number of agents. One thing that I notice in your plots is that the agents are not taking the same actions, which suggests they have different states/observations. In our evaluations in that paper, the agents start with an orientation aligned toward the goal. In the circle case, this means all agents should have the same observations at each time step.

from cadrl_ros.

We generate these 10 agents evenly across the circle, and they were all oriented toward their own goal. The initial positions were calculated, so the positions could be slightly different in their value, which will lead to slightly different decision-making outcomes.

The difference in decision making will further widen the difference in states, which ultimately leads to what shown in my plots.

from cadrl_ros.