{
'cameraEyePosition': [-1.0, -0.25, 0.6],
'cameraTargetPosition': [-0.6, -0.05, 0.2],
'cameraUpVector': [0, 0, 1],
'render_width': 224,
'render_height': 224
},
]
env = pmg.make_env(
task='pick_and_place',
gripper='parallel_jaw',
render=True,
binary_reward=True,
max_episode_steps=5,
image_observation=True,
depth_image=False,
goal_image=True,
visualize_target=True,
camera_setup=camera_setup,
observation_cam_id=[0],
goal_cam_id=1,
)
obs = env.reset()
t = 0
while True:
t += 1
action = env.action_space.sample()
obs, reward, done, info = env.step(action)
axarr[0].imshow(obs['desired_goal_img'])
axarr[1].imshow(obs['achieved_goal_img'])
plt.pause(0.00001)
'cameraEyePosition': [-0.9, -0.0, 0.4],
'cameraTargetPosition': [-0.45, -0.0, 0.0],
'cameraUpVector': [0, 0, 1],
'render_width': 224,
'render_height': 224
}
]
env = pmg.make_env(task='insertion',
render=True,
binary_reward=True,
distance_threshold=0.05,
image_observation=True,
depth_image=True,
goal_image=True,
point_cloud=False,
state_noise=True,
visualize_target=False,
camera_setup=camera_setup,
observation_cam_id=[1],
goal_cam_id=-1,
gripper='parallel_jaw',
max_episode_steps=50)
obs = env.reset()
time_done = False
f, axarr = plt.subplots(2, 2)
while True:
action = env.action_space.sample()
obs, reward, time_done, info = env.step(action)
axarr[0][0].imshow(obs['desired_goal_img'][:, :, :3])
'cameraUpVector': [0, 0, 1],
'render_width': 224,
'render_height': 224
},
]
env = pmg.make_env(
# task args
task='block_stack',
gripper='parallel_jaw',
grip_informed_goal=False,
num_block=5, # only meaningful for multi-block tasks
render=True,
binary_reward=True,
max_episode_steps=25,
# image observation args
image_observation=True,
depth_image=False,
goal_image=True,
visualize_target=True,
camera_setup=camera_setup,
observation_cam_id=[0],
goal_cam_id=1,
# task decomposition
task_decomposition=True)
"""The desired goal changes as the subgoal is being set"""
obs = env.reset()
time_done = False
env.set_sub_goal(0)
t = 0
path = os.path.dirname(os.path.realpath(__file__))
directory_name = args['task'] + '_' + str(args['num_blocks'])
params['curriculum'] = args['crcl']
if args['crcl']:
directory_name += '_crcl'
path = os.path.join(path, directory_name)
seed_returns = []
seed_success_rates = []
for seed in seeds:
# make env instance
env = pmg.make_env(task=args['task'],
gripper='parallel_jaw',
num_block=args['num_blocks'],
render=args['render'],
binary_reward=True,
image_observation=False,
use_curriculum=args['crcl'],
num_goals_to_generate=num_total_episodes,
max_episode_steps=max_episode_steps)
seed_path = path + '/seed' + str(seed)
agent = GoalConditionedDDPG(algo_params=params,
env=env,
path=seed_path,
seed=seed)
agent.run(test=False)
seed_returns.append(agent.statistic_dict['epoch_test_return'])
seed_success_rates.append(
agent.statistic_dict['epoch_test_success_rate'])
del env, agent
env_mujoco.reset()
done_mujoco = False
while not done_mujoco:
action = env_mujoco.action_space.sample()
_, _, done_mujoco, _ = env_mujoco.step(action)
cost = timeit.default_timer() - start_mujoco
costs_mujoco.append(cost / num_episodes)
print("Seed {}, mujoco average runtime over 100 episodes: {}".format(
seed, costs_mujoco[-1]))
# pmg loop
env_pmg = pmg.make_env(task='reach',
gripper='parallel_jaw',
render=False,
binary_reward=True,
max_episode_steps=50,
image_observation=False,
depth_image=False,
goal_image=False)
env_pmg.seed(seed)
start_pmg = timeit.default_timer()
for i in range(num_episodes):
env_pmg.reset()
done_pmg = False
while not done_pmg:
action = env_pmg.action_space.sample()
_, _, done_pmg, _ = env_pmg.step(action)
cost = timeit.default_timer() - start_pmg
costs_pmg.append(cost / num_episodes)
'cameraUpVector': [0, 0, 1],
'render_width': 224,
'render_height': 224
},
]
env = pmg.make_env(
# task args
task='block_rearrange',
gripper='parallel_jaw',
grip_informed_goal=False,
num_block=4, # only meaningful for multi-block tasks
render=True,
binary_reward=True,
max_episode_steps=5,
# image observation args
image_observation=True,
depth_image=False,
goal_image=True,
visualize_target=True,
camera_setup=camera_setup,
observation_cam_id=[0],
goal_cam_id=1,
# curriculum args
use_curriculum=True,
num_goals_to_generate=20)
"""You can expect the desired goal to change every once a while based on the current curriculum level,
and settle down at the hardest one finally."""
obs = env.reset()
env.activate_curriculum_update()
time_done = False
# directory for storing data
path = os.path.dirname(os.path.realpath(__file__))
directory_name = args['task']
if args['joint_ctrl']:
directory_name += '_joint_ctrl'
if args['hindsight']:
directory_name += '_her'
path = os.path.join(path, directory_name)
seed_returns = []
seed_success_rates = []
for seed in seeds:
# make env instance
env = pmg.make_env(task=args['task'],
gripper='parallel_jaw',
joint_control=args['joint_ctrl'],
render=args['render'],
binary_reward=True,
max_episode_steps=50)
seed_path = path + '/seed' + str(seed)
agent = GoalConditionedDDPG(algo_params=params,
env=env,
path=seed_path,
seed=seed)
agent.run(test=False)
seed_returns.append(agent.statistic_dict['epoch_test_return'])
seed_success_rates.append(
agent.statistic_dict['epoch_test_success_rate'])
del env, agent
import os
import numpy as np
import time
import pybullet_multigoal_gym as pmg
num_episodes = 32
env = pmg.make_env(task='block_rearrange',
gripper='parallel_jaw',
grip_informed_goal=False,
num_block=4,
render=True,
visualize_target=True,
binary_reward=True,
joint_control=False,
max_episode_steps=10000,
image_observation=False,
use_curriculum=True,
task_decomposition=False,
num_goals_to_generate=num_episodes)
env.activate_curriculum_update()
obs = env.reset(test=False)
time_done = False
while not time_done:
# time.sleep(0.1)
# action = env.action_space.sample() * 0
# action[-1] = -1
# obs, reward, time_done, info = env.step(action)
# if time_done:
# env.reset(test=False)
# time_done = False
