def test_bayrn_power(ex_dir, env: SimEnv, bayrn_hparam: dict):
pyrado.set_seed(0)
# Environments and domain randomization
env_real = deepcopy(env)
env_sim = DomainRandWrapperLive(env, create_zero_var_randomizer(env))
dp_map = create_default_domain_param_map_qq()
env_sim = MetaDomainRandWrapper(env_sim, dp_map)
env_real.domain_param = dict(mass_pend_pole=0.024 * 1.1,
mass_rot_pole=0.095 * 1.1)
env_real = wrap_like_other_env(env_real, env_sim)
# Policy and subroutine
policy_hparam = dict(energy_gain=0.587, ref_energy=0.827)
policy = QQubeSwingUpAndBalanceCtrl(env_sim.spec, **policy_hparam)
subrtn_hparam = dict(
max_iter=1,
pop_size=8,
num_init_states_per_domain=1,
num_is_samples=4,
expl_std_init=0.1,
num_workers=1,
)
subrtn = PoWER(ex_dir, env_sim, policy, **subrtn_hparam)
# Set the boundaries for the GP
dp_nom = inner_env(env_sim).get_nominal_domain_param()
ddp_space = BoxSpace(
bound_lo=np.array([
0.8 * dp_nom["mass_pend_pole"], 1e-8,
0.8 * dp_nom["mass_rot_pole"], 1e-8
]),
bound_up=np.array([
1.2 * dp_nom["mass_pend_pole"], 1e-7,
1.2 * dp_nom["mass_rot_pole"], 1e-7
]),
)
# Create algorithm and train
algo = BayRn(ex_dir,
env_sim,
env_real,
subrtn,
ddp_space,
**bayrn_hparam,
num_workers=1)
algo.train()
assert algo.curr_iter == algo.max_iter or algo.stopping_criterion_met()
# Experiment (set seed before creating the modules)
ex_dir = setup_experiment(
QQubeSwingUpSim.name,
f"{BayRn.name}-{PPO.name}_{FNNPolicy.name}",
"rand-mass_pend_pole-mass_rot_pole-length_pend_pole-length_rot_pole_lower-std",
)
# Set seed if desired
pyrado.set_seed(args.seed, verbose=True)
# Environments
env_sim_hparams = dict(dt=1 / 100.0, max_steps=600)
env_sim = QQubeSwingUpSim(**env_sim_hparams)
env_sim = ActNormWrapper(env_sim)
env_sim = DomainRandWrapperLive(env_sim,
create_zero_var_randomizer(env_sim))
dp_map = create_default_domain_param_map_qq()
env_sim = MetaDomainRandWrapper(env_sim, dp_map)
env_real_hparams = dict(dt=1 / 500.0, max_steps=3000)
env_real = QQubeSwingUpReal(**env_real_hparams)
env_real = wrap_like_other_env(env_real, env_sim)
# Policy
policy_hparam = dict(hidden_sizes=[64, 64], hidden_nonlin=to.tanh)
policy = FNNPolicy(spec=env_sim.spec, **policy_hparam)
# Critic
vfcn_hparam = dict(hidden_sizes=[32, 32], hidden_nonlin=to.tanh)
vfcn = FNNPolicy(spec=EnvSpec(env_sim.obs_space, ValueFunctionSpace),
**vfcn_hparam)
if __name__ == '__main__':
# Parse command line arguments
args = get_argparser().parse_args()
# Experiment (set seed before creating the modules)
ex_dir = setup_experiment(QQubeSwingUpSim.name,
f'{BayRn.name}-{PoWER.name}_{QQubeSwingUpAndBalanceCtrl.name}_sim2sim',
f'rand-Mp-Mr_seed-{args.seed}')
# Set seed if desired
pyrado.set_seed(args.seed, verbose=True)
# Environments
env_sim_hparams = dict(dt=1/100., max_steps=600)
env_sim = QQubeSwingUpSim(**env_sim_hparams)
env_sim = DomainRandWrapperLive(env_sim, create_zero_var_randomizer(env_sim))
dp_map = get_default_domain_param_map_qq()
env_sim = MetaDomainRandWrapper(env_sim, dp_map)
env_real = QQubeSwingUpSim(**env_sim_hparams)
env_real.domain_param = dict(
Mp=0.024*1.1,
Mr=0.095*1.1,
)
env_real_hparams = env_sim_hparams
env_real = wrap_like_other_env(env_real, env_sim)
# PoWER + energy-based controller setup
policy_hparam = dict(energy_gain=0.587, ref_energy=0.827, acc_max=10.)
policy = QQubeSwingUpAndBalanceCtrl(env_sim.spec, **policy_hparam)
subrtn_hparam = dict(