Skip to content
Branch: master
Find file Copy path
Find file Copy path
Fetching contributors…
Cannot retrieve contributors at this time
125 lines (100 sloc) 4.77 KB
#!/usr/bin/env python3
import gym
import numpy as np
from gym.envs.mujoco.mujoco_env import MujocoEnv
from learn2learn.gym.envs.meta_env import MetaEnv
def mass_center(model, sim):
mass = np.expand_dims(model.body_mass, 1)
xpos =
return (np.sum(mass * xpos, 0) / np.sum(mass))
class HumanoidDirectionEnv(MetaEnv, MujocoEnv, gym.utils.EzPickle):
This environment requires the humanoid to learn to run in a random direction in the
XY plane. At each time step the humanoid receives a signal composed of a
control cost and a reward equal to its average velocity in the target direction.
The tasks are 2d-arrays sampled uniformly along the unit circle.
The target direction is indicated by the vector from the origin to the sampled point.
The velocity is calculated as the distance (in the target direction) of the humanoid's torso
position before and after taking the specified action divided by a small value dt.
A small positive bonus is added to the reward to stop the humanoid from
prematurely ending the episode.
Adapted from Jonas Rothfuss' implementation.
1. Finn et al. 2017. "Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks." arXiv [cs.LG].
2. Rothfuss et al. 2018. "ProMP: Proximal Meta-Policy Search." arXiv [cs.LG].
def __init__(self, task=None):
MetaEnv.__init__(self, task)
MujocoEnv.__init__(self, 'humanoid.xml', 5)
# -------- MetaEnv Methods --------
def set_task(self, task):
MetaEnv.set_task(self, task)
self.goal_direction = task['direction']
def sample_tasks(self, num_tasks):
directions = np.random.normal(size=(num_tasks, 2))
directions /= np.linalg.norm(directions, axis=1)[..., np.newaxis]
tasks = [{'direction': direction} for direction in directions]
return tasks
# -------- Mujoco Methods --------
def _get_obs(self):
data =
return np.concatenate([data.qpos.flat[2:],
def viewer_setup(self): = 1 = self.model.stat.extent * 1.0 = -20
def reset_model(self):
c = 0.01
self.init_qpos + self.np_random.uniform(low=-c, high=c, size=self.model.nq),
self.init_qvel + self.np_random.uniform(low=-c, high=c, size=self.model.nv, )
return self._get_obs()
# -------- Gym Methods --------
def step(self, action):
pos_before = np.copy(mass_center(self.model, self.sim)[:2])
self.do_simulation(action, self.frame_skip)
pos_after = mass_center(self.model, self.sim)[:2]
alive_bonus = 5.0
data =
lin_vel_cost = 0.25 * np.sum(self.goal_direction * (pos_after - pos_before)) / self.model.opt.timestep
quad_ctrl_cost = 0.1 * np.square(data.ctrl).sum()
quad_impact_cost = .5e-6 * np.square(data.cfrc_ext).sum()
quad_impact_cost = min(quad_impact_cost, 10)
reward = lin_vel_cost - quad_ctrl_cost - quad_impact_cost + alive_bonus
qpos =
done = bool((qpos[2] < 1.0) or (qpos[2] > 2.0))
return self._get_obs(), reward, done, dict(reward_linvel=lin_vel_cost,
def reset(self, *args, **kwargs):
MujocoEnv.reset(self, *args, **kwargs)
return self._get_obs()
def render(self, mode='human'):
if mode == 'rgb_array':
# window size used for old mujoco-py:
width, height = 500, 500
data = self._get_viewer(mode).read_pixels(width,
return data
elif mode == 'human':
if __name__ == '__main__':
env = HumanoidDirectionEnv()
for task in [env.get_task(), env.sample_tasks(1)[0]]:
action = env.action_space.sample()
You can’t perform that action at this time.