Add MP-Wrapper for new TT variants with random init robot position
This commit is contained in:
parent
9faaae4785
commit
ad2f20a557
@ -29,6 +29,7 @@ from .mujoco.table_tennis.table_tennis_env import TableTennisEnv, TableTennisWin
|
|||||||
MAX_EPISODE_STEPS_TABLE_TENNIS, MAX_EPISODE_STEPS_TABLE_TENNIS_MARKOV_VER
|
MAX_EPISODE_STEPS_TABLE_TENNIS, MAX_EPISODE_STEPS_TABLE_TENNIS_MARKOV_VER
|
||||||
from .mujoco.table_tennis.mp_wrapper import TT_MPWrapper as MPWrapper_TableTennis
|
from .mujoco.table_tennis.mp_wrapper import TT_MPWrapper as MPWrapper_TableTennis
|
||||||
from .mujoco.table_tennis.mp_wrapper import TT_MPWrapper_Replan as MPWrapper_TableTennis_Replan
|
from .mujoco.table_tennis.mp_wrapper import TT_MPWrapper_Replan as MPWrapper_TableTennis_Replan
|
||||||
|
from .mujoco.table_tennis.mp_wrapper import TTRndRobot_MPWrapper as MPWrapper_TableTennis_Rnd
|
||||||
from .mujoco.table_tennis.mp_wrapper import TTVelObs_MPWrapper as MPWrapper_TableTennis_VelObs
|
from .mujoco.table_tennis.mp_wrapper import TTVelObs_MPWrapper as MPWrapper_TableTennis_VelObs
|
||||||
from .mujoco.table_tennis.mp_wrapper import TTVelObs_MPWrapper_Replan as MPWrapper_TableTennis_VelObs_Replan
|
from .mujoco.table_tennis.mp_wrapper import TTVelObs_MPWrapper_Replan as MPWrapper_TableTennis_VelObs_Replan
|
||||||
|
|
||||||
@ -306,6 +307,7 @@ register(
|
|||||||
register(
|
register(
|
||||||
id='fancy/TableTennisRndRobot-v0',
|
id='fancy/TableTennisRndRobot-v0',
|
||||||
entry_point='fancy_gym.envs.mujoco:TableTennisRandomInit',
|
entry_point='fancy_gym.envs.mujoco:TableTennisRandomInit',
|
||||||
|
mp_wrapper=MPWrapper_TableTennis_Rnd,
|
||||||
max_episode_steps=MAX_EPISODE_STEPS_TABLE_TENNIS,
|
max_episode_steps=MAX_EPISODE_STEPS_TABLE_TENNIS,
|
||||||
kwargs={
|
kwargs={
|
||||||
'random_pos_scale': 0.1,
|
'random_pos_scale': 0.1,
|
||||||
@ -315,6 +317,7 @@ register(
|
|||||||
|
|
||||||
register(
|
register(
|
||||||
id='fancy/TableTennisMarkovian-v0',
|
id='fancy/TableTennisMarkovian-v0',
|
||||||
|
mp_wrapper=MPWrapper_TableTennis,
|
||||||
entry_point='fancy_gym.envs.mujoco:TableTennisMarkovian',
|
entry_point='fancy_gym.envs.mujoco:TableTennisMarkovian',
|
||||||
max_episode_steps=MAX_EPISODE_STEPS_TABLE_TENNIS_MARKOV_VER,
|
max_episode_steps=MAX_EPISODE_STEPS_TABLE_TENNIS_MARKOV_VER,
|
||||||
kwargs={
|
kwargs={
|
||||||
@ -323,6 +326,7 @@ register(
|
|||||||
|
|
||||||
register(
|
register(
|
||||||
id='fancy/TableTennisRndRobotMarkovian-v0',
|
id='fancy/TableTennisRndRobotMarkovian-v0',
|
||||||
|
mp_wrapper=MPWrapper_TableTennis_Rnd,
|
||||||
entry_point='fancy_gym.envs.mujoco:TableTennisMarkovian',
|
entry_point='fancy_gym.envs.mujoco:TableTennisMarkovian',
|
||||||
max_episode_steps=MAX_EPISODE_STEPS_TABLE_TENNIS_MARKOV_VER,
|
max_episode_steps=MAX_EPISODE_STEPS_TABLE_TENNIS_MARKOV_VER,
|
||||||
kwargs={
|
kwargs={
|
||||||
|
@ -151,3 +151,15 @@ class TTVelObs_MPWrapper_Replan(TT_MPWrapper_Replan):
|
|||||||
[True] * 2, # target landing position
|
[True] * 2, # target landing position
|
||||||
# [True] * 1, # time
|
# [True] * 1, # time
|
||||||
])
|
])
|
||||||
|
|
||||||
|
class TTRndRobot_MPWrapper(TT_MPWrapper):
|
||||||
|
@property
|
||||||
|
def context_mask(self):
|
||||||
|
return np.hstack([
|
||||||
|
[True] * 7, # joints position
|
||||||
|
[False] * 7, # joints velocity
|
||||||
|
[True] * 2, # position ball x, y
|
||||||
|
[False] * 1, # position ball z
|
||||||
|
[True] * 2, # target landing position
|
||||||
|
# [True] * 1, # time
|
||||||
|
])
|
Loading…
Reference in New Issue
Block a user