dodox/fancy_gym
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

started cleaning up init. DMP envs are still not transferred. Wrappers for various environments still missing

Browse Source
This commit is contained in:
Onur
2022-06-28 20:33:19 +02:00
parent 9ad6fbe712
commit 8fe6a83271
3 changed files with 418 additions and 920 deletions
Download Patch File Download Diff File Expand all files Collapse all files
alr_envs/alr/mujoco/beerpong/beerpong.py
+31 -33
View File
@@ -178,8 +178,6 @@ class ALRBeerBongEnv(MujocoEnv, utils.EzPickle):
[self._steps],
])
def compute_reward(self):
@property
def dt(self):
return super(ALRBeerBongEnv, self).dt * self.repeat_action
@@ -213,37 +211,37 @@ class ALRBeerBongEnvStepBasedEpisodicReward(ALRBeerBongEnv):
return ob, reward, done, infos
class ALRBeerBongEnvStepBased(ALRBeerBongEnv):
def __init__(self, frame_skip=1, apply_gravity_comp=True, noisy=False, rndm_goal=False, cup_goal_pos=None):
super().__init__(frame_skip, apply_gravity_comp, noisy, rndm_goal, cup_goal_pos)
self.release_step = 62 # empirically evaluated for frame_skip=2!
def step(self, a):
if self._steps < self.release_step:
return super(ALRBeerBongEnvStepBased, self).step(a)
else:
reward = 0
done = False
while not done:
sub_ob, sub_reward, done, sub_infos = super(ALRBeerBongEnvStepBased, self).step(np.zeros(a.shape))
if not done or sub_infos['sim_crash']:
reward += sub_reward
else:
ball_pos = self.sim.data.body_xpos[self.sim.model._body_name2id["ball"]].copy()
cup_goal_dist_final = np.linalg.norm(ball_pos - self.sim.data.site_xpos[
self.sim.model._site_name2id["cup_goal_final_table"]].copy())
cup_goal_dist_top = np.linalg.norm(ball_pos - self.sim.data.site_xpos[
self.sim.model._site_name2id["cup_goal_table"]].copy())
if sub_infos['success']:
dist_rew = -cup_goal_dist_final ** 2
else:
dist_rew = -0.5 * cup_goal_dist_final ** 2 - cup_goal_dist_top ** 2
reward = reward - sub_infos['action_cost'] + dist_rew
infos = sub_infos
ob = sub_ob
ob[-1] = self.release_step + 1 # Since we simulate until the end of the episode, PPO does not see the
# internal steps and thus, the observation also needs to be set correctly
return ob, reward, done, infos
# class ALRBeerBongEnvStepBased(ALRBeerBongEnv):
# def __init__(self, frame_skip=1, apply_gravity_comp=True, noisy=False, rndm_goal=False, cup_goal_pos=None):
# super().__init__(frame_skip, apply_gravity_comp, noisy, rndm_goal, cup_goal_pos)
# self.release_step = 62 # empirically evaluated for frame_skip=2!
#
# def step(self, a):
# if self._steps < self.release_step:
# return super(ALRBeerBongEnvStepBased, self).step(a)
# else:
# reward = 0
# done = False
# while not done:
# sub_ob, sub_reward, done, sub_infos = super(ALRBeerBongEnvStepBased, self).step(np.zeros(a.shape))
# if not done or sub_infos['sim_crash']:
# reward += sub_reward
# else:
# ball_pos = self.sim.data.body_xpos[self.sim.model._body_name2id["ball"]].copy()
# cup_goal_dist_final = np.linalg.norm(ball_pos - self.sim.data.site_xpos[
# self.sim.model._site_name2id["cup_goal_final_table"]].copy())
# cup_goal_dist_top = np.linalg.norm(ball_pos - self.sim.data.site_xpos[
# self.sim.model._site_name2id["cup_goal_table"]].copy())
# if sub_infos['success']:
# dist_rew = -cup_goal_dist_final ** 2
# else:
# dist_rew = -0.5 * cup_goal_dist_final ** 2 - cup_goal_dist_top ** 2
# reward = reward - sub_infos['action_cost'] + dist_rew
# infos = sub_infos
# ob = sub_ob
# ob[-1] = self.release_step + 1 # Since we simulate until the end of the episode, PPO does not see the
# # internal steps and thus, the observation also needs to be set correctly
# return ob, reward, done, infos
if __name__ == "__main__":