def experiment_wo_distruber(env_real: RealEnv, env_sim: SimEnv):
# Wrap the environment in the same as done during training
env_real = wrap_like_other_env(env_real, env_sim)
# Run learned policy on the device
print_cbt('Running the evaluation policy ...', 'c', bright=True)
return rollout(env_real,
policy,
eval=True,
max_steps=args.max_steps,
render_mode=RenderMode(text=True),
no_reset=True,
no_close=True)
from pyrado.utils.argparser import get_argparser
if __name__ == '__main__':
# Parse command line arguments
args = get_argparser().parse_args()
# Get the experiment's directory to load from
ex_dir = ask_for_experiment()
# Load the policy (trained in simulation) and the environment (for constructing the real-world counterpart)
env_sim, policy, _ = load_experiment(ex_dir)
# Detect the correct real-world counterpart and create it
if isinstance(inner_env(env_sim), WAMBallInCupSim):
# If `max_steps` (or `dt`) are not explicitly set using `args`, use the same as in the simulation
max_steps = args.max_steps if args.max_steps < pyrado.inf else env_sim.max_steps
dt = args.dt if args.dt is not None else env_sim.dt
env_real = WAMBallInCupReal(dt=dt, max_steps=max_steps)
else:
raise pyrado.TypeErr(given=env_sim, expected_type=WAMBallInCupSim)
# Finally wrap the env in the same as done during training
env_real = wrap_like_other_env(env_real, env_sim)
# Run on device
done = False
while not done:
ro = rollout(env_real, policy, eval=True)
print_cbt(f'Return: {ro.undiscounted_return()}', 'g', bright=True)
done, _, _ = after_rollout_query(env_real, policy, ro)
def experiment_w_distruber(env_real: RealEnv, env_sim: SimEnv):
# Wrap the environment in the same as done during training
env_real = wrap_like_other_env(env_real, env_sim)
# Run learned policy on the device
print_cbt('Running the evaluation policy ...', 'c', bright=True)
ro1 = rollout(env_real,
policy,
eval=True,
max_steps=args.max_steps // 3,
render_mode=RenderMode(),
no_reset=True,
no_close=True)
# Run disturber
env_real = inner_env(env_real) # since we are reusing it
print_cbt('Running the 1st disturber ...', 'c', bright=True)
rollout(env_real,
disturber_pos,
eval=True,
max_steps=steps_disturb,
render_mode=RenderMode(),
no_reset=True,
no_close=True)
# Wrap the environment in the same as done during training
env_real = wrap_like_other_env(env_real, env_sim)
# Run learned policy on the device
print_cbt('Running the evaluation policy ...', 'c', bright=True)
ro2 = rollout(env_real,
policy,
eval=True,
max_steps=args.max_steps // 3,
render_mode=RenderMode(),
no_reset=True,
no_close=True)
# Run disturber
env_real = inner_env(env_real) # since we are reusing it
print_cbt('Running the 2nd disturber ...', 'c', bright=True)
rollout(env_real,
disturber_neg,
eval=True,
max_steps=steps_disturb,
render_mode=RenderMode(),
no_reset=True,
no_close=True)
# Wrap the environment in the same as done during training
env_real = wrap_like_other_env(env_real, env_sim)
# Run learned policy on the device
print_cbt('Running the evaluation policy ...', 'c', bright=True)
ro3 = rollout(env_real,
policy,
eval=True,
max_steps=args.max_steps // 3,
render_mode=RenderMode(),
no_reset=True,
no_close=True)
return StepSequence.concat([ro1, ro2, ro3])
df = pd.DataFrame(columns=['policy', 'ret', 'len'])
# Evaluate all policies
for i, (env_sim, policy) in enumerate(zip(env_sim_list, policy_list)):
# Create a new sampler pool for every policy to synchronize the random seeds i.e. init states
pool = SamplerPool(args.num_workers)
# Seed the sampler
if args.seed is not None:
pool.set_seed(args.seed)
print_cbt(f"Set the random number generators' seed to {args.seed}.", 'w')
else:
print_cbt('No seed was set', 'y')
# Add the same wrappers as during training
env = wrap_like_other_env(env, env_sim)
# Sample rollouts
ros = eval_randomized_domain(pool, env, pert, policy, init_state_list) # internally calls DomainRandWrapperLive
# Compute results metrics
rets = [ro.undiscounted_return() for ro in ros]
lengths = [float(ro.length) for ro in ros] # int values are not numeric in pandas
df = df.append(pd.DataFrame(dict(policy=ex_labels[i], ret=rets, len=lengths)), ignore_index=True)
metrics = dict(
avg_len=df.groupby('policy').mean()['len'].to_dict(),
avg_ret=df.groupby('policy').mean()['ret'].to_dict(),
median_ret=df.groupby('policy').median()['ret'].to_dict(),
min_ret=df.groupby('policy').min()['ret'].to_dict(),
max_ret=df.groupby('policy').max()['ret'].to_dict(),