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

added MPEnv

Browse Source
This commit is contained in:
ottofabian
2021-05-12 09:52:25 +02:00
parent b4ad3e6ddd
commit 95e9b8be47
11 changed files with 200 additions and 168 deletions
Download Patch File Download Diff File Expand all files Collapse all files
alr_envs/utils/legacy/utils.py
+5 -5
View File
@@ -1,7 +1,7 @@
import alr_envs.classic_control.hole_reacher as hr
import alr_envs.classic_control.viapoint_reacher as vpr
from alr_envs.utils.wrapper.dmp_wrapper import DmpWrapper
from alr_envs.utils.wrapper.detpmp_wrapper import DetPMPWrapper
from alr_envs.utils.mps.dmp_wrapper import DmpWrapper
from alr_envs.utils.mps.detpmp_wrapper import DetPMPWrapper
import numpy as np
@@ -65,7 +65,7 @@ def make_holereacher_env(rank, seed=0):
dt=_env.dt,
learn_goal=True,
alpha_phase=2,
start_pos=_env.start_pos,
start_pos=_env._start_pos,
policy_type="velocity",
weights_scale=50,
goal_scale=0.1
@@ -105,7 +105,7 @@ def make_holereacher_fix_goal_env(rank, seed=0):
learn_goal=False,
final_pos=np.array([2.02669572, -1.25966385, -1.51618198, -0.80946476, 0.02012344]),
alpha_phase=2,
start_pos=_env.start_pos,
start_pos=_env._start_pos,
policy_type="velocity",
weights_scale=50,
goal_scale=1
@@ -142,7 +142,7 @@ def make_holereacher_env_pmp(rank, seed=0):
num_basis=5,
width=0.02,
policy_type="velocity",
start_pos=_env.start_pos,
start_pos=_env._start_pos,
duration=2,
post_traj_time=0,
dt=_env.dt,
alr_envs/utils/make_env_helpers.py
+2 -2
View File
@@ -1,5 +1,5 @@
from alr_envs.utils.wrapper.dmp_wrapper import DmpWrapper
from alr_envs.utils.wrapper.detpmp_wrapper import DetPMPWrapper
from alr_envs.utils.mps.dmp_wrapper import DmpWrapper
from alr_envs.utils.mps.detpmp_wrapper import DetPMPWrapper
import gym
from gym.vector.utils import write_to_shared_memory
import sys
alr_envs/utils/wrapper/__init__.py → alr_envs/utils/mps/__init__.py
View File
alr_envs/utils/wrapper/detpmp_wrapper.py → alr_envs/utils/mps/detpmp_wrapper.py
+7 -6
View File
@@ -2,17 +2,18 @@ import gym
import numpy as np
from mp_lib import det_promp
from alr_envs.utils.wrapper.mp_wrapper import MPWrapper
from alr_envs.utils.mps.mp_environments import MPEnv
from alr_envs.utils.mps.mp_wrapper import MPWrapper
class DetPMPWrapper(MPWrapper):
def __init__(self, env, num_dof, num_basis, width, start_pos=None, duration=1, dt=0.01, post_traj_time=0.,
policy_type=None, weights_scale=1, zero_start=False, zero_goal=False, **mp_kwargs):
def __init__(self, env: MPEnv, num_dof: int, num_basis: int, width: int, start_pos=None, duration: int = 1,
dt: float = 0.01, post_traj_time: float = 0., policy_type: str = None, weights_scale: float = 1.,
zero_start: bool = False, zero_goal: bool = False, **mp_kwargs):
# self.duration = duration # seconds
super().__init__(env, num_dof, duration, dt, post_traj_time, policy_type, weights_scale,
num_basis=num_basis, width=width, start_pos=start_pos, zero_start=zero_start,
zero_goal=zero_goal)
super().__init__(env, num_dof, dt, duration, post_traj_time, policy_type, weights_scale, num_basis=num_basis,
width=width, start_pos=start_pos, zero_start=zero_start, zero_goal=zero_goal, **mp_kwargs)
action_bounds = np.inf * np.ones((self.mp.n_basis * self.mp.n_dof))
self.action_space = gym.spaces.Box(low=-action_bounds, high=action_bounds, dtype=np.float32)
alr_envs/utils/wrapper/dmp_wrapper.py → alr_envs/utils/mps/dmp_wrapper.py
+14 -42
View File
@@ -1,19 +1,18 @@
from mp_lib.phase import ExpDecayPhaseGenerator
from mp_lib.basis import DMPBasisGenerator
from mp_lib import dmps
import numpy as np
import gym
import numpy as np
from mp_lib import dmps
from mp_lib.basis import DMPBasisGenerator
from mp_lib.phase import ExpDecayPhaseGenerator
from alr_envs.utils.wrapper.mp_wrapper import MPWrapper
from alr_envs.utils.mps.mp_environments import MPEnv
from alr_envs.utils.mps.mp_wrapper import MPWrapper
class DmpWrapper(MPWrapper):
def __init__(self, env: gym.Env, num_dof: int, num_basis: int,
# start_pos: np.ndarray = None,
# final_pos: np.ndarray = None,
def __init__(self, env: MPEnv, num_dof: int, num_basis: int,
duration: int = 1, alpha_phase: float = 2., dt: float = None,
learn_goal: bool = False, return_to_start: bool = False, post_traj_time: float = 0.,
learn_goal: bool = False, post_traj_time: float = 0.,
weights_scale: float = 1., goal_scale: float = 1., bandwidth_factor: float = 3.,
policy_type: str = None, render_mode: str = None):
@@ -23,8 +22,6 @@ class DmpWrapper(MPWrapper):
env:
num_dof:
num_basis:
start_pos:
final_pos:
duration:
alpha_phase:
dt:
@@ -37,30 +34,17 @@ class DmpWrapper(MPWrapper):
self.learn_goal = learn_goal
dt = env.dt if hasattr(env, "dt") else dt
assert dt is not None
# start_pos = start_pos if start_pos is not None else env.start_pos if hasattr(env, "start_pos") else None
# TODO: assert start_pos is not None # start_pos will be set in initialize, do we need this here?
# if learn_goal:
# final_pos = np.zeros_like(start_pos) # arbitrary, will be learned
# final_pos = np.zeros((1, num_dof)) # arbitrary, will be learned
# else:
# final_pos = final_pos if final_pos is not None else start_pos if return_to_start else None
# assert final_pos is not None
self.t = np.linspace(0, duration, int(duration / dt))
self.goal_scale = goal_scale
super().__init__(env, num_dof, duration, dt, post_traj_time, policy_type, weights_scale, render_mode,
num_basis=num_basis,
# start_pos=start_pos, final_pos=final_pos,
alpha_phase=alpha_phase,
bandwidth_factor=bandwidth_factor)
super().__init__(env, num_dof, dt, duration, post_traj_time, policy_type, weights_scale, render_mode,
num_basis=num_basis, alpha_phase=alpha_phase, bandwidth_factor=bandwidth_factor)
action_bounds = np.inf * np.ones((np.prod(self.mp.dmp_weights.shape) + (num_dof if learn_goal else 0)))
self.action_space = gym.spaces.Box(low=-action_bounds, high=action_bounds, dtype=np.float32)
def initialize_mp(self, num_dof: int, duration: int, dt: float, num_basis: int = 5,
# start_pos: np.ndarray = None,
# final_pos: np.ndarray = None,
alpha_phase: float = 2., bandwidth_factor: float = 3.):
def initialize_mp(self, num_dof: int, duration: int, dt: float, num_basis: int = 5, alpha_phase: float = 2.,
bandwidth_factor: int = 3):
phase_generator = ExpDecayPhaseGenerator(alpha_phase=alpha_phase, duration=duration)
basis_generator = DMPBasisGenerator(phase_generator, duration=duration, num_basis=num_basis,
@@ -69,15 +53,6 @@ class DmpWrapper(MPWrapper):
dmp = dmps.DMP(num_dof=num_dof, basis_generator=basis_generator, phase_generator=phase_generator,
num_time_steps=int(duration / dt), dt=dt)
# dmp.dmp_start_pos = start_pos.reshape((1, num_dof))
# in a contextual environment, the start_pos may be not fixed, set in mp_rollout?
# TODO: Should we set start_pos in init at all? It's only used after calling rollout anyway...
# dmp.dmp_start_pos = start_pos.reshape((1, num_dof)) if start_pos is not None else np.zeros((1, num_dof))
# weights = np.zeros((num_basis, num_dof))
# goal_pos = np.zeros(num_dof) if self.learn_goal else final_pos
# dmp.set_weights(weights, goal_pos)
return dmp
def goal_and_weights(self, params):
@@ -87,18 +62,15 @@ class DmpWrapper(MPWrapper):
if self.learn_goal:
goal_pos = params[0, -self.mp.num_dimensions:] # [num_dof]
params = params[:, :-self.mp.num_dimensions] # [1,num_dof]
# weight_matrix = np.reshape(params[:, :-self.num_dof], [self.num_basis, self.num_dof])
else:
goal_pos = self.env.goal_pos # self.mp.dmp_goal_pos.flatten()
assert goal_pos is not None
# weight_matrix = np.reshape(params, [self.num_basis, self.num_dof])
weight_matrix = np.reshape(params, self.mp.dmp_weights.shape)
weight_matrix = np.reshape(params, self.mp.dmp_weights.shape) # [num_basis, num_dof]
return goal_pos * self.goal_scale, weight_matrix * self.weights_scale
def mp_rollout(self, action):
# if self.mp.start_pos is None:
self.mp.dmp_start_pos = self.env.init_qpos # start_pos
self.mp.dmp_start_pos = self.env.start_pos
goal_pos, weight_matrix = self.goal_and_weights(action)
self.mp.set_weights(weight_matrix, goal_pos)
return self.mp.reference_trajectory(self.t)
alr_envs/utils/mps/mp_environments.py
+33
View File
@@ -0,0 +1,33 @@
from abc import abstractmethod
from typing import Union
import gym
import numpy as np
class MPEnv(gym.Env):
@property
@abstractmethod
def corrected_obs_index(self):
"""Returns boolean value for each observation entry
whether the observation is returned by the DMP for the contextual case or not.
This effectively allows to filter unwanted or unnecessary observations from the full step-based case.
"""
raise NotImplementedError()
@property
@abstractmethod
def start_pos(self) -> Union[float, int, np.ndarray]:
"""
Returns the current position of the joints
"""
raise NotImplementedError()
@property
def goal_pos(self) -> Union[float, int, np.ndarray]:
"""
Returns the current final position of the joints for the MP.
By default this returns the starting position.
"""
return self.start_pos
alr_envs/utils/wrapper/mp_wrapper.py → alr_envs/utils/mps/mp_wrapper.py
+10 -39
View File
@@ -1,32 +1,18 @@
from abc import ABC, abstractmethod
from collections import defaultdict
import gym
import numpy as np
from alr_envs.utils.mps.mp_environments import MPEnv
from alr_envs.utils.policies import get_policy_class
class MPWrapper(gym.Wrapper, ABC):
def __init__(self,
env: gym.Env,
num_dof: int,
duration: int = 1,
dt: float = None,
post_traj_time: float = 0.,
policy_type: str = None,
weights_scale: float = 1.,
render_mode: str = None,
**mp_kwargs
):
def __init__(self, env: MPEnv, num_dof: int, dt: float, duration: int = 1, post_traj_time: float = 0.,
policy_type: str = None, weights_scale: float = 1., render_mode: str = None, **mp_kwargs):
super().__init__(env)
# self.num_dof = num_dof
# self.num_basis = num_basis
# self.duration = duration # seconds
# dt = env.dt if hasattr(env, "dt") else dt
assert dt is not None # this should never happen as MPWrapper is a base class
self.post_traj_steps = int(post_traj_time / dt)
@@ -40,8 +26,11 @@ class MPWrapper(gym.Wrapper, ABC):
self.render_mode = render_mode
self.render_kwargs = {}
# TODO: not yet final
# TODO: @Max I think this should not be in this class, this functionality should be part of your sampler.
def __call__(self, params, contexts=None):
"""
Can be used to provide a batch of parameter sets
"""
params = np.atleast_2d(params)
obs = []
rewards = []
@@ -63,7 +52,7 @@ class MPWrapper(gym.Wrapper, ABC):
def reset(self):
obs = self.env.reset()
return obs
return obs[self.env]
def step(self, action: np.ndarray):
""" This function generates a trajectory based on a DMP and then does the usual loop over reset and step"""
@@ -77,15 +66,9 @@ class MPWrapper(gym.Wrapper, ABC):
# self._velocity = velocity
rewards = 0
# infos = defaultdict(list)
# TODO: @Max Why do we need this configure, states should be part of the model
# TODO: Ask Onur if the context distribution needs to be outside the environment
# TODO: For now create a new env with each context
# TODO: Explicitly call reset before step to obtain context from obs?
# self.env.configure(context)
# obs = self.env.reset()
info = {}
# create random obs as the reset function is called externally
obs = self.env.observation_space.sample()
for t, pos_vel in enumerate(zip(trajectory, velocity)):
ac = self.policy.get_action(pos_vel[0], pos_vel[1])
@@ -107,18 +90,6 @@ class MPWrapper(gym.Wrapper, ABC):
self.render_mode = mode
self.render_kwargs = kwargs
# def __call__(self, actions):
# return self.step(actions)
# params = np.atleast_2d(params)
# rewards = []
# infos = []
# for p, c in zip(params, contexts):
# reward, info = self.rollout(p, c)
# rewards.append(reward)
# infos.append(info)
#
# return np.array(rewards), infos
@abstractmethod
def mp_rollout(self, action):
"""