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

merge branches

Browse Source
This commit is contained in:
Maximilian Huettenrauch
2021-04-19 11:53:30 +02:00
parent 51e8503873 dee2fad263
commit f3d75b9a60
23 changed files with 796 additions and 622 deletions
Download Patch File Download Diff File Expand all files Collapse all files
alr_envs/mujoco/__init__.py
+2 -1
View File
@@ -1 +1,2 @@
from alr_envs.mujoco.reacher.alr_reacher import ALRReacherEnv
from alr_envs.mujoco.reacher.alr_reacher import ALRReacherEnv
from alr_envs.mujoco.balancing import BalancingEnv
alr_envs/mujoco/balancing.py
+53
View File
@@ -0,0 +1,53 @@
import os
import numpy as np
from gym import utils
from gym.envs.mujoco import mujoco_env
from alr_envs.utils.utils import angle_normalize
class BalancingEnv(mujoco_env.MujocoEnv, utils.EzPickle):
def __init__(self, n_links=5):
utils.EzPickle.__init__(**locals())
self.n_links = n_links
if n_links == 5:
file_name = 'reacher_5links.xml'
elif n_links == 7:
file_name = 'reacher_7links.xml'
else:
raise ValueError(f"Invalid number of links {n_links}, only 5 or 7 allowed.")
mujoco_env.MujocoEnv.__init__(self, os.path.join(os.path.dirname(__file__), "assets", file_name), 2)
def step(self, a):
angle = angle_normalize(np.sum(self.sim.data.qpos.flat[:self.n_links]), type="rad")
reward = - np.abs(angle)
self.do_simulation(a, self.frame_skip)
ob = self._get_obs()
done = False
return ob, reward, done, dict(angle=angle, end_effector=self.get_body_com("fingertip").copy())
def viewer_setup(self):
self.viewer.cam.trackbodyid = 1
def reset_model(self):
# This also generates a goal, we however do not need/use it
qpos = self.np_random.uniform(low=-0.1, high=0.1, size=self.model.nq) + self.init_qpos
qpos[-2:] = 0
qvel = self.init_qvel + self.np_random.uniform(low=-.005, high=.005, size=self.model.nv)
qvel[-2:] = 0
self.set_state(qpos, qvel)
return self._get_obs()
def _get_obs(self):
theta = self.sim.data.qpos.flat[:self.n_links]
return np.concatenate([
np.cos(theta),
np.sin(theta),
self.sim.data.qvel.flat[:self.n_links], # this is angular velocity
])
alr_envs/mujoco/ball_in_a_cup/utils.py
+56 -56
View File
@@ -1,5 +1,5 @@
from alr_envs.utils.detpmp_env_wrapper import DetPMPEnvWrapper
from alr_envs.utils.dmp_env_wrapper import DmpEnvWrapper
from alr_envs.utils.wrapper.detpmp_wrapper import DetPMPWrapper
from alr_envs.utils.wrapper.dmp_wrapper import DmpWrapper
from alr_envs.mujoco.ball_in_a_cup.ball_in_a_cup import ALRBallInACupEnv
from alr_envs.mujoco.ball_in_a_cup.ball_in_a_cup_simple import ALRBallInACupEnv as ALRBallInACupEnvSimple
@@ -18,19 +18,19 @@ def make_contextual_env(rank, seed=0):
def _init():
env = ALRBallInACupEnv()
env = DetPMPEnvWrapper(env,
num_dof=7,
num_basis=5,
width=0.005,
policy_type="motor",
start_pos=env.start_pos,
duration=3.5,
post_traj_time=4.5,
dt=env.dt,
weights_scale=0.5,
zero_start=True,
zero_goal=True
)
env = DetPMPWrapper(env,
num_dof=7,
num_basis=5,
width=0.005,
policy_type="motor",
start_pos=env.start_pos,
duration=3.5,
post_traj_time=4.5,
dt=env.dt,
weights_scale=0.5,
zero_start=True,
zero_goal=True
)
env.seed(seed + rank)
return env
@@ -52,19 +52,19 @@ def make_env(rank, seed=0):
def _init():
env = ALRBallInACupEnvSimple()
env = DetPMPEnvWrapper(env,
num_dof=7,
num_basis=5,
width=0.005,
policy_type="motor",
start_pos=env.start_pos,
duration=3.5,
post_traj_time=4.5,
dt=env.dt,
weights_scale=0.2,
zero_start=True,
zero_goal=True
)
env = DetPMPWrapper(env,
num_dof=7,
num_basis=5,
width=0.005,
policy_type="motor",
start_pos=env.start_pos,
duration=3.5,
post_traj_time=4.5,
dt=env.dt,
weights_scale=0.2,
zero_start=True,
zero_goal=True
)
env.seed(seed + rank)
return env
@@ -86,20 +86,20 @@ def make_simple_env(rank, seed=0):
def _init():
env = ALRBallInACupEnvSimple()
env = DetPMPEnvWrapper(env,
num_dof=3,
num_basis=5,
width=0.005,
off=-0.1,
policy_type="motor",
start_pos=env.start_pos[1::2],
duration=3.5,
post_traj_time=4.5,
dt=env.dt,
weights_scale=0.25,
zero_start=True,
zero_goal=True
)
env = DetPMPWrapper(env,
num_dof=3,
num_basis=5,
width=0.005,
off=-0.1,
policy_type="motor",
start_pos=env.start_pos[1::2],
duration=3.5,
post_traj_time=4.5,
dt=env.dt,
weights_scale=0.25,
zero_start=True,
zero_goal=True
)
env.seed(seed + rank)
return env
@@ -121,20 +121,20 @@ def make_simple_dmp_env(rank, seed=0):
def _init():
_env = ALRBallInACupEnvSimple()
_env = DmpEnvWrapper(_env,
num_dof=3,
num_basis=5,
duration=3.5,
post_traj_time=4.5,
bandwidth_factor=2.5,
dt=_env.dt,
learn_goal=False,
alpha_phase=3,
start_pos=_env.start_pos[1::2],
final_pos=_env.start_pos[1::2],
policy_type="motor",
weights_scale=100,
)
_env = DmpWrapper(_env,
num_dof=3,
num_basis=5,
duration=3.5,
post_traj_time=4.5,
bandwidth_factor=2.5,
dt=_env.dt,
learn_goal=False,
alpha_phase=3,
start_pos=_env.start_pos[1::2],
final_pos=_env.start_pos[1::2],
policy_type="motor",
weights_scale=100,
)
_env.seed(seed + rank)
return _env
alr_envs/mujoco/beerpong/utils.py
+40 -40
View File
@@ -1,4 +1,4 @@
from alr_envs.utils.detpmp_env_wrapper import DetPMPEnvWrapper
from alr_envs.utils.wrapper.detpmp_wrapper import DetPMPWrapper
from alr_envs.mujoco.beerpong.beerpong import ALRBeerpongEnv
from alr_envs.mujoco.beerpong.beerpong_simple import ALRBeerpongEnv as ALRBeerpongEnvSimple
@@ -17,19 +17,19 @@ def make_contextual_env(rank, seed=0):
def _init():
env = ALRBeerpongEnv()
env = DetPMPEnvWrapper(env,
num_dof=7,
num_basis=5,
width=0.005,
policy_type="motor",
start_pos=env.start_pos,
duration=3.5,
post_traj_time=4.5,
dt=env.dt,
weights_scale=0.5,
zero_start=True,
zero_goal=True
)
env = DetPMPWrapper(env,
num_dof=7,
num_basis=5,
width=0.005,
policy_type="motor",
start_pos=env.start_pos,
duration=3.5,
post_traj_time=4.5,
dt=env.dt,
weights_scale=0.5,
zero_start=True,
zero_goal=True
)
env.seed(seed + rank)
return env
@@ -51,19 +51,19 @@ def make_env(rank, seed=0):
def _init():
env = ALRBeerpongEnvSimple()
env = DetPMPEnvWrapper(env,
num_dof=7,
num_basis=5,
width=0.005,
policy_type="motor",
start_pos=env.start_pos,
duration=3.5,
post_traj_time=4.5,
dt=env.dt,
weights_scale=0.25,
zero_start=True,
zero_goal=True
)
env = DetPMPWrapper(env,
num_dof=7,
num_basis=5,
width=0.005,
policy_type="motor",
start_pos=env.start_pos,
duration=3.5,
post_traj_time=4.5,
dt=env.dt,
weights_scale=0.25,
zero_start=True,
zero_goal=True
)
env.seed(seed + rank)
return env
@@ -85,19 +85,19 @@ def make_simple_env(rank, seed=0):
def _init():
env = ALRBeerpongEnvSimple()
env = DetPMPEnvWrapper(env,
num_dof=3,
num_basis=5,
width=0.005,
policy_type="motor",
start_pos=env.start_pos[1::2],
duration=3.5,
post_traj_time=4.5,
dt=env.dt,
weights_scale=0.5,
zero_start=True,
zero_goal=True
)
env = DetPMPWrapper(env,
num_dof=3,
num_basis=5,
width=0.005,
policy_type="motor",
start_pos=env.start_pos[1::2],
duration=3.5,
post_traj_time=4.5,
dt=env.dt,
weights_scale=0.5,
zero_start=True,
zero_goal=True
)
env.seed(seed + rank)
return env
alr_envs/mujoco/reacher/alr_reacher.py
+10 -25
View File
@@ -9,6 +9,8 @@ from alr_envs.utils.utils import angle_normalize
class ALRReacherEnv(mujoco_env.MujocoEnv, utils.EzPickle):
def __init__(self, steps_before_reward=200, n_links=5, balance=False):
utils.EzPickle.__init__(**locals())
self._steps = 0
self.steps_before_reward = steps_before_reward
self.n_links = n_links
@@ -29,65 +31,48 @@ class ALRReacherEnv(mujoco_env.MujocoEnv, utils.EzPickle):
else:
raise ValueError(f"Invalid number of links {n_links}, only 5 or 7 allowed.")
self._q_pos = []
self._q_vel = []
utils.EzPickle.__init__(self)
mujoco_env.MujocoEnv.__init__(self, os.path.join(os.path.dirname(__file__), "assets", file_name), 2)
@property
def current_pos(self):
return self.sim.data.qpos[0:5].copy()
@property
def current_vel(self):
return self.sim.data.qvel[0:5].copy()
def step(self, a):
self._steps += 1
reward_dist = 0.0
angular_vel = 0.0
reward_balance = 0.0
if self._steps >= self.steps_before_reward:
vec = self.get_body_com("fingertip") - self.get_body_com("target")
reward_dist -= self.reward_weight * np.linalg.norm(vec)
angular_vel -= np.linalg.norm(self.sim.data.qvel.flat[:self.n_links])
reward_ctrl = - np.square(a).sum()
reward_balance = - self.balance_weight * np.abs(
angle_normalize(np.sum(self.sim.data.qpos.flat[:self.n_links]), type="rad"))
if self.balance:
reward_balance -= self.balance_weight * np.abs(
angle_normalize(np.sum(self.sim.data.qpos.flat[:self.n_links]), type="rad"))
reward = reward_dist + reward_ctrl + angular_vel + reward_balance
self.do_simulation(a, self.frame_skip)
self._q_pos.append(self.sim.data.qpos[0:5].ravel().copy())
self._q_vel.append(self.sim.data.qvel[0:5].ravel().copy())
ob = self._get_obs()
done = False
return ob, reward, done, dict(reward_dist=reward_dist, reward_ctrl=reward_ctrl,
velocity=angular_vel, reward_balance=reward_balance,
end_effector=self.get_body_com("fingertip").copy(),
goal=self.goal if hasattr(self, "goal") else None,
traj=self._q_pos, vel=self._q_vel)
goal=self.goal if hasattr(self, "goal") else None)
def viewer_setup(self):
self.viewer.cam.trackbodyid = 0
def reset_model(self):
qpos = self.init_qpos # self.np_random.uniform(low=-0.1, high=0.1, size=self.model.nq) + self.init_qpos
qpos = self.np_random.uniform(low=-0.1, high=0.1, size=self.model.nq) + self.init_qpos
while True:
self.goal = self.np_random.uniform(low=-self.n_links / 10, high=self.n_links / 10, size=2)
if np.linalg.norm(self.goal) < self.n_links / 10:
break
qpos[-2:] = self.goal
qvel = self.init_qvel # + self.np_random.uniform(low=-.005, high=.005, size=self.model.nv)
qvel = self.init_qvel + self.np_random.uniform(low=-.005, high=.005, size=self.model.nv)
qvel[-2:] = 0
self.set_state(qpos, qvel)
self._steps = 0
self._q_pos = []
self._q_vel = []
return self._get_obs()
def _get_obs(self):