def _player_step_tuple(self, envs_step_tuples):
"""Construct observation, return usual step tuple.
Args:
envs_step_tuples: tuples.
Returns:
Step tuple: ob, reward, done, info
ob: concatenated images [simulated observation, real observation,
difference], with additional informations in header.
reward: real environment reward
done: True iff. envs_step_tuples['real_env'][2] is True
info: real environment info
"""
ob_real, reward_real, _, _ = envs_step_tuples["real_env"]
ob_sim, reward_sim, _, _ = envs_step_tuples["sim_env"]
ob_err = absolute_hinge_difference(ob_sim, ob_real)
ob_real_aug = self._augment_observation(ob_real, reward_real,
self.cumulative_real_reward)
ob_sim_aug = self._augment_observation(ob_sim, reward_sim,
self.cumulative_sim_reward)
ob_err_aug = self._augment_observation(
ob_err, reward_sim - reward_real,
self.cumulative_sim_reward - self.cumulative_real_reward)
ob = np.concatenate([ob_sim_aug, ob_real_aug, ob_err_aug], axis=1)
_, reward, done, info = envs_step_tuples["real_env"]
return ob, reward, done, info
def append_debug_frame_batch(sim_obs, real_obs, sim_cum_rews,
real_cum_rews, sim_rews, real_rews):
"""Add a debug frame."""
rews = [[sim_cum_rews, sim_rews], [real_cum_rews, real_rews]]
headers = []
for j in range(len(sim_obs)):
local_nps = []
for i in range(2):
img = PIL_Image().new(
"RGB",
(sim_obs.shape[-2], 11),
)
draw = PIL_ImageDraw().Draw(img)
draw.text((0, 0),
"c:{:3}, r:{:3}".format(int(rews[i][0][j]),
int(rews[i][1][j])),
fill=(255, 0, 0))
local_nps.append(np.asarray(img))
local_nps.append(np.zeros_like(local_nps[0]))
headers.append(np.concatenate(local_nps, axis=1))
errs = absolute_hinge_difference(sim_obs, real_obs)
headers = np.stack(headers)
debug_frame_batches.append( # pylint: disable=cell-var-from-loop
np.concatenate([
headers,
np.concatenate([sim_obs, real_obs, errs], axis=2)
],
axis=1))
