Continous control agent in plane with obstacles
wmkouw opened this issue · comments
In this wonderful paper by Huebotter et al., they demonstrated an agent in a 2D planar environment that navigates to a setpoint, which, upon the agent reaching it, jumps to a new location. It's a bit like the old mobile game, Snake, but then in a continuous state-space with a continuous action-space.
I propose we build a similar agent but also add obstacles to this plane. We can use chance constraints (see also #68) to indicate that the part of state-space the obstacle occupies, should not be traversed. The agent should find a trajectory around it.
The next step would be to learn the parameters of the chance constraints through interactions, i.e., the agent bumps into the obstacle and incurs a large prediction error.
Chengfeng implemented a ship control agent operating on a 2d plane, see folder ship.
Implementing obstacles, through for instance chance constraints, is still a todo.
Hey @wmkouw @Chengfeng-Jia! Any developments on this one? I see the ETA is on May 18.
Considering that Bart has already utilized NUV prior to making control decisions to avoid obstacles, I suggest perhaps closing this issue.