Embedded Files
Effects of other Interaction Parameters
Effects of other Interaction Parameters
Some examples showing effects of different parameters on behaviors.
Some examples showing effects of different parameters on behaviors.
Effect of different Speed limit: assign different speed limits to the green agent.
Green : max v = 4 m/sec
Green : max v = 2 m/sec
Effect of different state cost weight Q: assign different state cost weight to the blue agent.
Big Q : more aggressive
Small Q : less aggressive
Blue : Q = [0.0005, 0.0005, 0, 0]
Blue : Q = [0.001, 0.001, 0, 0]
Observation field and occlusion
Observation field and occlusion
One important factor in interaction is whether the agents are observed by other agents in the interaction. As described in the paper, we only consider a potential game happening when other agents are not occluded and in the observation range.
One important factor in interaction is whether the agents are observed by other agents in the interaction. As described in the paper, we only consider a potential game happening when other agents are not occluded and in the observation range.
- Observation Range
The agent is visible if it is in the front or back visible range and not blind by the obstacles.
The agent is visible if it is in the front or back visible range and not blind by the obstacles.
Blind by obstacles
Do not blind by obstacles
Front & Back visible range
The red triangle region indicates the observed interaction.
Blind by obstacles
Front & Back : has the furthest visable distance
Observing is not "symmetric"!
Observing is not "symmetric"!
- I saw you, but did you see me?
- I saw you, but did you see me?
- How can I estimate if other agents can see me?(future work!)
- How can I estimate if other agents can see me?(future work!)
In the right case,:
In the right case,:
- Purple and Red observe each other so IPG exists for both two.
- Orange observes Blue but Blue does not observe Orange. So Blue does not yield to Orange and deadlock occurs.
Deadlock solution
Deadlock solution
One of the main causes to deadlocks in the distributed settings is that each agent assumes that the other will make the first move (taking two-agent scenarios for example).
One of the main causes to deadlocks in the distributed settings is that each agent assumes that the other will make the first move (taking two-agent scenarios for example).
To mitigate this issue, we propose a method that involves adjusting the estimated goal points online (to current position) when deadlock happens to promote yielding behavior. The results below demonstrate the effectiveness of this approach in the hallway scenario.
To mitigate this issue, we propose a method that involves adjusting the estimated goal points online (to current position) when deadlock happens to promote yielding behavior. The results below demonstrate the effectiveness of this approach in the hallway scenario.
Speed: x2.5
Speed: x2.5
Page updated
Google Sites
Report abuse