import os
from osim.env import OsimEnv
## Define an environment where we teach an agent to stand still
# We use a walker model and define the objective to keep the center of mass still
class StandingEnv(OsimEnv):
model_path = os.path.join(os.path.dirname(__file__), '../osim/models/gait9dof18musc.osim')
time_limit = 300
ninput = 99
def is_done(self):
# End the simulation if the pelvis is too low
state_desc = self.get_state_desc()
return state_desc["joint_pos"]["ground_pelvis"][2] < 0.7
def get_observation(self):
state_desc = self.get_state_desc()
# Augmented environment from the L2R challenge
res = []
# Map some of the state variables to the observation vector
for body_part in ["pelvis","head","torso","toes_l","toes_r","talus_l","talus_r"]:
res = res + state_desc["body_pos_rot"][body_part][2:]
res = res + state_desc["body_pos"][body_part][0:2]
res = res + state_desc["body_vel_rot"][body_part][2:]
res = res + state_desc["body_vel"][body_part][0:2]
res = res + state_desc["body_acc_rot"][body_part][2:]
res = res + state_desc["body_acc"][body_part][0:2]
for joint in ["ankle_l","ankle_r","back","ground_pelvis","hip_l","hip_r","knee_l","knee_r"]:
res = res + state_desc["joint_pos"][joint]
res = res + state_desc["joint_vel"][joint]
res = res + state_desc["joint_acc"][joint]
res = res + state_desc["misc"]["mass_center_pos"] + state_desc["misc"]["mass_center_vel"] + state_desc["misc"]["mass_center_acc"]
return res
def get_observation_space_size(self):
return 99
def reward(self):
# Get the current state and the last state
state_desc = self.get_state_desc()
prev_state_desc = self.get_prev_state_desc()
if not prev_state_desc:
return 0
# Penalize movement of the pelvis
res = -(prev_state_desc["misc"]["mass_center_pos"][0] - state_desc["misc"]["mass_center_pos"][0])**2\
-(prev_state_desc["misc"]["mass_center_pos"][1] - state_desc["misc"]["mass_center_pos"][1])**2
# Penalize very low position of the pelvis
res += -(state_desc["joint_pos"]["ground_pelvis"][2] < 0.8)
return res
env = StandingEnv(visualize=True)
observation = env.reset()
for i in range(200):
observation, reward, done, info = env.step(env.action_space.sample())
print("Reward %f" % reward)
if done:
env.reset()
Datasets
Models
