added MPEnv
This commit is contained in:
ottofabian
@@ -0,0 +1,41 @@
|
||||
import gym
|
||||
import numpy as np
|
||||
from mp_lib import det_promp
|
||||
|
||||
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: 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, 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)
|
||||
|
||||
self.start_pos = start_pos
|
||||
self.dt = dt
|
||||
|
||||
def initialize_mp(self, num_dof: int, duration: int, dt: float, num_basis: int = 5, width: float = None,
|
||||
start_pos: np.ndarray = None, zero_start: bool = False, zero_goal: bool = False):
|
||||
pmp = det_promp.DeterministicProMP(n_basis=num_basis, n_dof=num_dof, width=width, off=0.01,
|
||||
zero_start=zero_start, zero_goal=zero_goal)
|
||||
|
||||
weights = np.zeros(shape=(num_basis, num_dof))
|
||||
pmp.set_weights(duration, weights)
|
||||
|
||||
return pmp
|
||||
|
||||
def mp_rollout(self, action):
|
||||
params = np.reshape(action, (self.mp.n_basis, self.mp.n_dof)) * self.weights_scale
|
||||
self.mp.set_weights(self.duration, params)
|
||||
_, des_pos, des_vel, _ = self.mp.compute_trajectory(1 / self.dt, 1.)
|
||||
if self.mp.zero_start:
|
||||
des_pos += self.start_pos[None, :]
|
||||
|
||||
return des_pos, des_vel
|
||||
@@ -0,0 +1,76 @@
|
||||
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.mps.mp_environments import MPEnv
|
||||
from alr_envs.utils.mps.mp_wrapper import MPWrapper
|
||||
|
||||
|
||||
class DmpWrapper(MPWrapper):
|
||||
|
||||
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, 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):
|
||||
|
||||
"""
|
||||
This Wrapper generates a trajectory based on a DMP and will only return episodic performances.
|
||||
Args:
|
||||
env:
|
||||
num_dof:
|
||||
num_basis:
|
||||
duration:
|
||||
alpha_phase:
|
||||
dt:
|
||||
learn_goal:
|
||||
post_traj_time:
|
||||
policy_type:
|
||||
weights_scale:
|
||||
goal_scale:
|
||||
"""
|
||||
self.learn_goal = learn_goal
|
||||
dt = env.dt if hasattr(env, "dt") else dt
|
||||
assert dt is not None
|
||||
self.t = np.linspace(0, duration, int(duration / dt))
|
||||
self.goal_scale = goal_scale
|
||||
|
||||
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, 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,
|
||||
basis_bandwidth_factor=bandwidth_factor)
|
||||
|
||||
dmp = dmps.DMP(num_dof=num_dof, basis_generator=basis_generator, phase_generator=phase_generator,
|
||||
num_time_steps=int(duration / dt), dt=dt)
|
||||
|
||||
return dmp
|
||||
|
||||
def goal_and_weights(self, params):
|
||||
assert params.shape[-1] == self.action_space.shape[0]
|
||||
params = np.atleast_2d(params)
|
||||
|
||||
if self.learn_goal:
|
||||
goal_pos = params[0, -self.mp.num_dimensions:] # [num_dof]
|
||||
params = params[:, :-self.mp.num_dimensions] # [1,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.mp.dmp_weights.shape) # [num_basis, num_dof]
|
||||
return goal_pos * self.goal_scale, weight_matrix * self.weights_scale
|
||||
|
||||
def mp_rollout(self, action):
|
||||
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)
|
||||
@@ -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
|
||||
@@ -0,0 +1,110 @@
|
||||
from abc import ABC, abstractmethod
|
||||
|
||||
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: 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)
|
||||
|
||||
assert dt is not None # this should never happen as MPWrapper is a base class
|
||||
self.post_traj_steps = int(post_traj_time / dt)
|
||||
|
||||
self.mp = self.initialize_mp(num_dof, duration, dt, **mp_kwargs)
|
||||
self.weights_scale = weights_scale
|
||||
|
||||
policy_class = get_policy_class(policy_type)
|
||||
self.policy = policy_class(env)
|
||||
|
||||
# rendering
|
||||
self.render_mode = render_mode
|
||||
self.render_kwargs = {}
|
||||
|
||||
# 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 = []
|
||||
dones = []
|
||||
infos = []
|
||||
# for p, c in zip(params, contexts):
|
||||
for p in params:
|
||||
# self.configure(c)
|
||||
ob, reward, done, info = self.step(p)
|
||||
obs.append(ob)
|
||||
rewards.append(reward)
|
||||
dones.append(done)
|
||||
infos.append(info)
|
||||
|
||||
return obs, np.array(rewards), dones, infos
|
||||
|
||||
def configure(self, context):
|
||||
self.env.configure(context)
|
||||
|
||||
def reset(self):
|
||||
obs = self.env.reset()
|
||||
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"""
|
||||
trajectory, velocity = self.mp_rollout(action)
|
||||
|
||||
if self.post_traj_steps > 0:
|
||||
trajectory = np.vstack([trajectory, np.tile(trajectory[-1, :], [self.post_traj_steps, 1])])
|
||||
velocity = np.vstack([velocity, np.zeros(shape=(self.post_traj_steps, self.mp.num_dimensions))])
|
||||
|
||||
# self._trajectory = trajectory
|
||||
# self._velocity = velocity
|
||||
|
||||
rewards = 0
|
||||
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])
|
||||
obs, rew, done, info = self.env.step(ac)
|
||||
rewards += rew
|
||||
# TODO return all dicts?
|
||||
# [infos[k].append(v) for k, v in info.items()]
|
||||
if self.render_mode:
|
||||
self.env.render(mode=self.render_mode, **self.render_kwargs)
|
||||
if done:
|
||||
break
|
||||
|
||||
done = True
|
||||
return obs, rewards, done, info
|
||||
|
||||
def render(self, mode='human', **kwargs):
|
||||
"""Only set render options here, such that they can be used during the rollout.
|
||||
This only needs to be called once"""
|
||||
self.render_mode = mode
|
||||
self.render_kwargs = kwargs
|
||||
|
||||
@abstractmethod
|
||||
def mp_rollout(self, action):
|
||||
"""
|
||||
Generate trajectory and velocity based on the MP
|
||||
Returns:
|
||||
trajectory/positions, velocity
|
||||
"""
|
||||
raise NotImplementedError()
|
||||
|
||||
@abstractmethod
|
||||
def initialize_mp(self, num_dof: int, duration: int, dt: float, **kwargs):
|
||||
"""
|
||||
Create respective instance of MP
|
||||
Returns:
|
||||
MP instance
|
||||
"""
|
||||
|
||||
raise NotImplementedError
|
||||
Reference in New Issue
Block a user