def __init__(self, rsc, cfg, visualize):
self.env = rsg_anymal.RaisimGymEnv(
rsc, dump(cfg['environment'], Dumper=RoundTripDumper), visualize)
self.num_acts = self.env.getActionDim()
self.num_obs = self.env.getObDim()
self.action_space = spaces.Box(np.ones(self.num_acts) * -1.,
np.ones(self.num_acts) * 1.,
dtype=np.float32)
self.observation_space = spaces.Box(np.ones(self.num_obs) * -100.,
np.ones(self.num_obs) * 100.,
dtype=np.float32)
self.reward_range = (-float('inf'), float('inf'))
self.metadata = {'render.modes': []}
home_path = task_path + "/../../../.."
#argument
test_mode = sys.argv[1] == 'True'
# config
cfg = YAML().load(open(task_path + "/cfg.yaml", 'r'))
curriculum_start = cfg['environment']['curriculum']['curriculum_start']
# create environment from the configuration file
if test_mode:
cfg_tmp = cfg
cfg_tmp['environment']['num_envs'] = 1
env = VecEnv(
rsg_anymal.RaisimGymEnv(
task_path + "/anymal",
dump(cfg_tmp['environment'], Dumper=RoundTripDumper)),
cfg['environment'])
else:
env = VecEnv(
rsg_anymal.RaisimGymEnv(
task_path + "/anymal",
dump(cfg['environment'], Dumper=RoundTripDumper)),
cfg['environment'])
# shortcuts
ob_dim = env.num_obs
act_dim = env.num_acts
# save the configuration and other files
saver = ConfigurationSaver(log_dir=home_path + "/data",
mode = args.mode
weight_path = args.weight
# check if gpu is available
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# directories
task_path = os.path.dirname(os.path.realpath(__file__))
home_path = task_path + "/../../../../.."
# config
cfg = YAML().load(open(task_path + "/cfg.yaml", 'r'))
# create environment from the configuration file
env = VecEnv(
rsg_anymal.RaisimGymEnv(home_path + "/rsc",
dump(cfg['environment'], Dumper=RoundTripDumper)),
cfg['environment'])
# shortcuts
ob_dim = env.num_obs
act_dim = env.num_acts
# Training
n_steps = math.floor(cfg['environment']['max_time'] /
cfg['environment']['control_dt'])
total_steps = n_steps * env.num_envs
avg_rewards = []
actor = ppo_module.Actor(
ppo_module.MLP(cfg['architecture']['policy_net'], nn.LeakyReLU, ob_dim,
import gym
import os
from ruamel.yaml import YAML, dump, RoundTripDumper
from stable_baselines3 import PPO
from stable_baselines3.ppo import MlpPolicy
from raisimGymTorch.env.bin import rsg_anymal
from raisimGymTorch.stable_baselines3.RaisimSbGymVecEnv import RaisimSbGymVecEnv as VecEnv
# Parallel environments
# directories
stb_path = os.path.dirname(os.path.realpath(__file__))
rsc_path = stb_path + "/../../../rsc"
task_path = stb_path + "/../env/envs/rsg_anymal"
# config
cfg = YAML().load(open(task_path + "/cfg.yaml", 'r'))
# create environment from the configuration file
env = VecEnv(rsg_anymal.RaisimGymEnv(rsc_path, dump(cfg['environment'], Dumper=RoundTripDumper)), cfg['environment'])
obs = env.reset()
n_steps = int(cfg['environment']['max_time'] / cfg['environment']['control_dt'])
model = PPO(MlpPolicy, env,
n_steps=n_steps,
verbose=1,
batch_size=int(n_steps*env.num_envs/4),
n_epochs=4)
model.learn(total_timesteps=250000000)
