def configure_ve_her(params):
env = cached_make_env(params['make_env'])
env.reset(reset_goal=False)
def reward_fun(ag_2, g, info): # vectorized
return env.compute_reward(achieved_goal=ag_2,
desired_goal=g,
info=info)
# Prepare configuration for HER.
ddpg_her_params = {
'reward_fun': reward_fun,
}
for name in ['replay_strategy', 'replay_k']:
ddpg_her_params[name] = params[name]
# params['_' + name] = ddpg_her_params[name]
del params[name]
params['ddpg_her_params'] = ddpg_her_params
ddpg_sample_transitions = make_sample_her_transitions(**ddpg_her_params)
ve_her_params = {
'reward_fun': reward_fun,
}
for name in ['replay_strategy', 'replay_k']:
ve_her_params[name] = params[f've_{name}']
# params[f'_ve_{name}'] = ve_her_params[name]
del params[f've_{name}']
params['ve_her_params'] = ve_her_params
ve_sample_transitions = make_sample_her_transitions(**ve_her_params)
return ddpg_sample_transitions, ve_sample_transitions
def configure_ve_her(params):
env = cached_make_env(params['make_env'])
# try:
# env.reset()
# except NotImplementedError:
# env.get_reset_obs()
def reward_fun(ag_2, g, info): # vectorized
return env.compute_reward(achieved_goal=ag_2, desired_goal=g, info=info)
# Prepare configuration for HER.
ddpg_her_params = {
'reward_fun': reward_fun,
}
for name in ['replay_strategy', 'replay_k']:
ddpg_her_params[name] = params[name]
params['_' + name] = ddpg_her_params[name]
del params[name]
ddpg_sample_transitions = make_sample_her_transitions(**ddpg_her_params)
ve_her_params = {
'reward_fun': reward_fun,
}
for name in ['replay_strategy', 'replay_k']:
ve_her_params[name] = params[f've_{name}']
params[f'_ve_{name}'] = ve_her_params[name]
del params[f've_{name}']
ve_sample_transitions = make_sample_her_transitions(**ve_her_params)
return ddpg_sample_transitions, ve_sample_transitions
def configure_her(params):
print("Inside configure her")
env = cached_make_env(params['make_env'])
env.reset()
def reward_fun(ag_2, g, info): # vectorized
return env.compute_reward(achieved_goal=ag_2,
desired_goal=g,
info=info)
# Prepare configuration for HER.
her_params = {
'reward_fun': reward_fun,
}
for name in ['replay_strategy', 'replay_k']:
her_params[name] = params[name]
params['_' + name] = her_params[name]
del params[name]
if params['prioritization'] == 'energy':
sample_her_transitions = make_sample_her_transitions_energy(
**her_params)
elif params['prioritization'] == 'tderror':
sample_her_transitions = make_sample_her_transitions_prioritized_replay(
**her_params)
else:
sample_her_transitions = make_sample_her_transitions(**her_params)
return sample_her_transitions
def configure_her(params):
try:
env = cached_make_env(params['make_env'])
env.reset()
def reward_fun(ag_2, g, info): # vectorized
return env.compute_reward(achieved_goal=ag_2, desired_goal=g, info=info)
except:
def reward_fun(ag_2, g, info): # vectorized
achieved_goal=ag_2
desired_goal=g
info=info
try:
d = np.linalg.norm(achieved_goal[:,0] - desired_goal[:,0], axis=-1)
fragile_goal = np.linalg.norm((achieved_goal[:,1:5] - desired_goal[:,1:5]), axis=-1) + np.linalg.norm((achieved_goal[:,5:] - desired_goal[:,5:])*((achieved_goal[:,5:] - desired_goal[:,5:]) < 0), axis=-1)
except:
d = np.linalg.norm(achieved_goal[0] - desired_goal[0], axis=-1)
fragile_goal = np.linalg.norm((achieved_goal[1:5] - desired_goal[1:5]), axis=-1) + np.linalg.norm((achieved_goal[5:] - desired_goal[5:])*((achieved_goal[5:] - desired_goal[5:]) < 0), axis=-1)
return -(d > np.pi/16).astype(np.float32) - np.float32(fragile_goal) * 2.0
# Prepare configuration for HER.
her_params = {
'reward_fun': reward_fun,
}
for name in ['replay_strategy', 'replay_k']:
her_params[name] = params[name]
params['_' + name] = her_params[name]
del params[name]
sample_her_transitions = make_sample_her_transitions(**her_params)
return sample_her_transitions
def configure_her(params):
env = cached_make_env(params['make_env'])
env.reset()
# goal_indexes = params['goal_indexes']
policy_index = params['policy_index']
#can add policy index here for subgoal rewards****
def reward_fun(ag_2, g, info): # vectorized
return env.compute_reward(achieved_goal=ag_2,
desired_goal=g,
info=info)
# Prepare configuration for HER.
her_params = {
'reward_fun': reward_fun,
# 'policy_indexes': goal_indexes,
'policy_index': policy_index,
}
for name in ['replay_strategy', 'replay_k']:
her_params[name] = params[name]
params['_' + name] = her_params[name]
del params[name]
sample_her_transitions = make_sample_her_transitions(**her_params)
return sample_her_transitions
def configure_her(params):
env = cached_make_env(params['make_env'])
env.reset()
def reward_fun(ag_2, g, info): # vectorized
return env.compute_reward(achieved_goal=ag_2, desired_goal=g, info=info)
# Prepare configuration for HER.
her_params = {
'reward_fun': reward_fun,
}
for name in ['replay_strategy', 'replay_k']:
her_params[name] = params[name]
params['_' + name] = her_params[name]
del params[name]
sample_her_transitions = make_sample_her_transitions(**her_params)
return sample_her_transitions
def configure_her(params):
env = cached_make_env(params['make_env'])
env.reset()
def reward_fun(ag_2, g, info): # vectorized
return env.compute_reward(achieved_goal=ag_2, desired_goal=g, info=info)
# Prepare configuration for HER.
her_params = {
'reward_fun': reward_fun,
}
for name in ['replay_strategy', 'replay_k']:
her_params[name] = params[name]
params['_' + name] = her_params[name]
del params[name]
sample_her_transitions = make_sample_her_transitions(**her_params)
return sample_her_transitions
def configure_her(params):
env = cached_make_env(params['make_env'])
env.reset()
def getMinObjOriGoal(curr_obs):
min_obj_ori_goal = 1000
max_obj_ori_goal = 0
for row in curr_obs:
rel_pos = row[6:9]
#print('rel_pos', rel_pos)
obj_ori_goal = rel_pos.copy()
obj_ori_goal[2] += 0.03
#print("obj_ori_goal", obj_ori_goal, obj_ori_goal[2])
if np.linalg.norm(obj_ori_goal) <= min_obj_ori_goal:
min_obj_ori_goal = np.linalg.norm(obj_ori_goal)
if np.linalg.norm(obj_ori_goal) >= max_obj_ori_goal:
max_obj_ori_goal = np.linalg.norm(obj_ori_goal)
#print("min val", min_obj_ori_goal)
#print("max val", max_obj_ori_goal)
return min_obj_ori_goal
def reward_fun(o, ag_2, g, info): # vectorized
global minDist
curr_obs = o.copy()
w = -3.0
#w = -5.0 #showed 0.9
min_obj_ori_goal = getMinObjOriGoal(curr_obs)
return w * min_obj_ori_goal + env.compute_reward(
achieved_goal=ag_2, desired_goal=g, info=info)
# Prepare configuration for HER.
her_params = {
'reward_fun': reward_fun,
}
for name in ['replay_strategy', 'replay_k']:
her_params[name] = params[name]
params['_' + name] = her_params[name]
del params[name]
sample_her_transitions = make_sample_her_transitions(**her_params)
return sample_her_transitions