You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The UdeM students have reported that the instructions to launch the simulator launch a different version of the simulator.
We can fix this by not running the server via Docker, but rather than throwing this code inside of the training scripts:
You will not want to start the Docker container as cited in the instructions, if you've already started it, please run: docker stop gym-duckietown-server.
Then, replacegym.make(...) in your training script with:
Note: Do not replace the gym.make(params['env']) inside of solution.py!
from gym_duckietown.simulator import Simulator
env = Simulator(
seed=123, # random seed
map_name="loop_empty",
max_steps=500001, # we don't want the gym to reset itself
domain_rand=0,
camera_width=640,
camera_height=480,
accept_start_angle_deg=4, # start close to straight
full_transparency=True,
distortion=True,
)
# Wrappers
env = ResizeWrapper(env)
env = NormalizeWrapper(env)
env = ImgWrapper(env) # to make the images from 160x120x3 into 3x160x120
env = ActionWrapper(env)
# env = DtRewardWrapper(env) # not during testing
and then, the output of the environment will match exactly what we have inside of the evaluator.
This is a hot fix while I work on Dockerizing this, but this should work for the students to get back on track.
I actually mean that the images were undistorted and the actions were different in the simulator when using the old instructions.
We still make sure on our end (i.e the evaluation) that we randomize the environments, so its not cheating, but this will give you the correct type of environment (i.e distorted / fish eye images + wheel velocities as actions).
The UdeM students have reported that the instructions to launch the simulator launch a different version of the simulator.
We can fix this by not running the server via Docker, but rather than throwing this code inside of the training scripts:
You will not want to start the Docker container as cited in the instructions, if you've already started it, please run:
docker stop gym-duckietown-server
.Then, replace
gym.make(...)
in your training script with:Note: Do not replace the
gym.make(params['env'])
inside ofsolution.py
!and then, the output of the environment will match exactly what we have inside of the evaluator.
This is a hot fix while I work on Dockerizing this, but this should work for the students to get back on track.
@liampaull @fgolemo
The text was updated successfully, but these errors were encountered: