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

added dmc2gym conversion and example how to leverage DMPs

Browse Source
This commit is contained in:
ottofabian
2021-06-28 17:25:53 +02:00
parent c8742e2934
commit 3b215cd877
17 changed files with 573 additions and 143 deletions
Download Patch File Download Diff File Expand all files Collapse all files
alr_envs/examples/__init__.py
View File
alr_envs/examples/examples_dmc.py
+73
View File
@@ -0,0 +1,73 @@
from alr_envs.dmc.Ball_in_the_cup_mp_wrapper import BallInCupMPWrapper
from alr_envs.utils.make_env_helpers import make_dmp_env, make_env
def example_dmc(env_name="fish-swim", seed=1):
env = make_env(env_name, seed)
rewards = 0
obs = env.reset()
# number of samples/full trajectories (multiple environment steps)
for i in range(2000):
ac = env.action_space.sample()
obs, reward, done, info = env.step(ac)
rewards += reward
if done:
print(rewards)
rewards = 0
obs = env.reset()
def example_custom_dmc_and_mp(seed=1):
"""
Example for running a custom motion primitive based environments based off of a dmc task.
Our already registered environments follow the same structure, but do not directly allow for modifications.
Hence, this also allows to adjust hyperparameters of the motion primitives more easily.
We appreciate PRs for custom environments (especially MP wrappers of existing tasks)
for our repo: https://github.com/ALRhub/alr_envs/
Args:
seed: seed
Returns:
"""
base_env = "ball_in_cup-catch"
# Replace this wrapper with the custom wrapper for your environment by inheriting from the MPEnvWrapper.
# You can also add other gym.Wrappers in case they are needed.
# wrappers = [HoleReacherMPWrapper]
wrappers = [BallInCupMPWrapper]
mp_kwargs = {
"num_dof": 2, # env.start_pos
"num_basis": 5,
"duration": 2,
"learn_goal": True,
"alpha_phase": 2,
"bandwidth_factor": 2,
"policy_type": "velocity",
"weights_scale": 50,
"goal_scale": 0.1
}
env = make_dmp_env(base_env, wrappers=wrappers, seed=seed, **mp_kwargs)
# OR for a deterministic ProMP:
# env = make_detpmp_env(base_env, wrappers=wrappers, seed=seed, **mp_args)
rewards = 0
obs = env.reset()
# number of samples/full trajectories (multiple environment steps)
for i in range(10):
ac = env.action_space.sample()
obs, reward, done, info = env.step(ac)
rewards += reward
if done:
print(rewards)
rewards = 0
obs = env.reset()
if __name__ == '__main__':
example_dmc()
example_custom_dmc_and_mp()
alr_envs/examples/examples_general.py
+74
View File
@@ -0,0 +1,74 @@
import warnings
from collections import defaultdict
import gym
import numpy as np
from alr_envs.utils.make_env_helpers import make_env
from alr_envs.utils.mp_env_async_sampler import AlrContextualMpEnvSampler, AlrMpEnvSampler, DummyDist
def example_general(env_id='alr_envs:ALRReacher-v0', seed=1):
"""
Example for running any env in the step based setting.
This also includes DMC environments when leveraging our custom make_env function.
"""
env = make_env(env_id, seed)
rewards = 0
obs = env.reset()
print("Observation shape: ", obs.shape)
print("Action shape: ", env.action_space.shape)
# number of environment steps
for i in range(10000):
obs, reward, done, info = env.step(env.action_space.sample())
rewards += reward
# if i % 1 == 0:
# env.render()
if done:
print(rewards)
rewards = 0
obs = env.reset()
def example_async(env_id="alr_envs:HoleReacherDMP-v0", n_cpu=4, seed=int('533D', 16)):
def sample(env: gym.vector.VectorEnv, n_samples=100):
# for plotting
rewards = np.zeros(n_cpu)
# this would generate more samples than requested if n_samples % num_envs != 0
repeat = int(np.ceil(n_samples / env.num_envs))
vals = defaultdict(list)
for i in range(repeat):
obs, reward, done, info = envs.step(envs.action_space.sample())
vals['obs'].append(obs)
vals['reward'].append(reward)
vals['done'].append(done)
vals['info'].append(info)
rewards += reward
if np.any(done):
print(rewards[done])
rewards[done] = 0
# do not return values above threshold
return (*map(lambda v: np.stack(v)[:n_samples], vals.values()),)
from alr_envs.utils.make_env_helpers import make_env_rank
envs = gym.vector.AsyncVectorEnv([make_env_rank(env_id, seed, i) for i in range(n_cpu)])
# envs = gym.vector.AsyncVectorEnv([make_env(env_id, seed + i) for i in range(n_cpu)])
obs = envs.reset()
print(sample(envs, 16))
if __name__ == '__main__':
# DMC
# example_general("fish-swim")
# custom mujoco env
# example_general("alr_envs:ALRReacher-v0")
example_general("ball_in_cup-catch")
alr_envs/examples/examples_motion_primitives.py
+103
View File
@@ -0,0 +1,103 @@
from alr_envs import HoleReacherMPWrapper
from alr_envs.utils.make_env_helpers import make_dmp_env, make_env
def example_mp(env_name="alr_envs:HoleReacherDMP-v1", seed=1):
"""
Example for running a motion primitive based environment, which is already registered
Args:
env_name: DMP env_id
seed: seed
Returns:
"""
# While in this case gym.make() is possible to use as well, we recommend our custom make env function.
# First, it already takes care of seeding and second enables the use of DMC tasks within the gym interface.
env = make_env(env_name, seed)
rewards = 0
# env.render(mode=None)
obs = env.reset()
# number of samples/full trajectories (multiple environment steps)
for i in range(10):
ac = env.action_space.sample()
obs, reward, done, info = env.step(ac)
rewards += reward
if i % 1 == 0:
# render full DMP trajectory
# render can only be called once in the beginning as well. That would render every trajectory
# Calling it after every trajectory allows to modify the mode. mode=None, disables rendering.
env.render(mode="human")
if done:
print(rewards)
rewards = 0
obs = env.reset()
def example_custom_mp(seed=1):
"""
Example for running a custom motion primitive based environments.
Our already registered environments follow the same structure, but do not directly allow for modifications.
Hence, this also allows to adjust hyperparameters of the motion primitives more easily.
We appreciate PRs for custom environments (especially MP wrappers of existing tasks)
for our repo: https://github.com/ALRhub/alr_envs/
Args:
seed: seed
Returns:
"""
base_env = "alr_envs:HoleReacher-v1"
# Replace this wrapper with the custom wrapper for your environment by inheriting from the MPEnvWrapper.
# You can also add other gym.Wrappers in case they are needed.
wrappers = [HoleReacherMPWrapper]
mp_kwargs = {
"num_dof": 5,
"num_basis": 5,
"duration": 2,
"learn_goal": True,
"alpha_phase": 2,
"bandwidth_factor": 2,
"policy_type": "velocity",
"weights_scale": 50,
"goal_scale": 0.1
}
env = make_dmp_env(base_env, wrappers=wrappers, seed=seed, **mp_kwargs)
# OR for a deterministic ProMP:
# env = make_detpmp_env(base_env, wrappers=wrappers, seed=seed)
rewards = 0
# env.render(mode=None)
obs = env.reset()
# number of samples/full trajectories (multiple environment steps)
for i in range(10):
ac = env.action_space.sample()
obs, reward, done, info = env.step(ac)
rewards += reward
if i % 1 == 0:
# render full DMP trajectory
# render can only be called once in the beginning as well. That would render every trajectory
# Calling it after every trajectory allows to modify the mode. mode=None, disables rendering.
env.render(mode="human")
if done:
print(rewards)
rewards = 0
obs = env.reset()
if __name__ == '__main__':
# DMP
example_mp("alr_envs:HoleReacherDMP-v1")
# DetProMP
example_mp("alr_envs:HoleReacherDetPMP-v1")
# Custom DMP
example_custom_mp()