Introducing a single agent, multi-goal, four-ways intersection environment #699
Conversation
This reverts commit d9bf4bb.
…ation. (I am not 100% sure if this is OK. Might cause error.)
pengzhenghao
changed the title
…disable shortcut config map.
===== timestep 220 ===== route completion: route_completion/goals/default: 0.92 route_completion/goals/go_straight: 0.50 route_completion/goals/left_turn: 0.46 route_completion/goals/right_turn: 0.92 route_completion/goals/u_turn: 0.53 reward: reward/default_reward: 1.27 reward/goal_agnostic_reward: 0.05 reward/goals/default: 1.27 reward/goals/go_straight: 0.14 reward/goals/left_turn: 0.06 reward/goals/right_turn: 1.27 reward/goals/u_turn: 0.01 =======================
# Conflicts: # metadrive/component/vehicle/base_vehicle.py
|
This PR is ready. Just merge after all tests passed. |
pengzhenghao
added
the
Merge after all tests pass
|
|
|
Identified that bug comes from:
In this PR, we build the "lane line" for "lanes in the negative road" too. I believe this should be the expected behavior and previous one which didn't build lane line for negative road was wrong. This affects the side detector and cause the observation mismatch in I don't think the "lane line in the negative road" will cause significant change to the system, and those value mismatch in test can be ignored by replacing the GT values by the new one. To verify that, this is the new side detector visualization in an arbitrary map by
This is the old side detector visualization:
|
What changes do you make in this PR?
This environment will maintain 4 goals: turn-left, turn-right, go-straight, u-turn and report the observation, termination, reward for each of these goals into the info dict. So a policy can be trained to conduct any of this four goals.
TODO:
Checklist
bash scripts/format.shbefore merging.