def observe(self) -> np.ndarray:
if not self.env.road:
return np.zeros(
(3, 3,
int(self.horizon * self.env.config["policy_frequency"])))
grid = compute_ttc_grid(self.env,
vehicle=self.observer_vehicle,
time_quantization=1 /
self.env.config["policy_frequency"],
horizon=self.horizon)
padding = np.ones(np.shape(grid))
padded_grid = np.concatenate([padding, grid, padding], axis=1)
obs_lanes = 3
l0 = grid.shape[1] + self.observer_vehicle.lane_index[
2] - obs_lanes // 2
lf = grid.shape[1] + self.observer_vehicle.lane_index[
2] + obs_lanes // 2
clamped_grid = padded_grid[:, l0:lf + 1, :]
repeats = np.ones(clamped_grid.shape[0])
repeats[np.array([0, -1])] += clamped_grid.shape[0]
padded_grid = np.repeat(clamped_grid, repeats.astype(int), axis=0)
obs_speeds = 3
v0 = grid.shape[0] + self.observer_vehicle.speed_index - obs_speeds // 2
vf = grid.shape[0] + self.observer_vehicle.speed_index + obs_speeds // 2
clamped_grid = padded_grid[v0:vf + 1, :, :]
return clamped_grid
def observe(self) -> np.ndarray:
if not self.env.road:
return np.zeros(self.space().shape)
# Add ego-vehicle
df = pd.DataFrame.from_records([self.observer_vehicle.to_dict()])[self.features]
# Add nearby traffic
# sort = self.order == "sorted"
close_vehicles = self.env.road.close_vehicles_to(self.observer_vehicle,
self.env.PERCEPTION_DISTANCE,
count=self.vehicles_count - 1,
see_behind=self.see_behind)
if close_vehicles:
origin = self.observer_vehicle if not self.absolute else None
df = df.append(pd.DataFrame.from_records(
[v.to_dict(origin, observe_intentions=self.observe_intentions)
for v in close_vehicles[-self.vehicles_count + 1:]])[self.features],
ignore_index=True)
# Normalize and clip
if self.normalize:
df = self.normalize_obs(df)
# Fill missing rows
if df.shape[0] < self.vehicles_count:
rows = np.zeros((self.vehicles_count - df.shape[0], len(self.features)))
df = df.append(pd.DataFrame(data=rows, columns=self.features), ignore_index=True)
# Reorder
df = df[self.features]
obs = df.values.copy()
if self.order == "shuffled":
self.env.np_random.shuffle(obs[1:])
# Flatten
if not self.env.road:
return np.zeros((3, 3, int(self.horizon * self.env.config["policy_frequency"])))
grid = compute_ttc_grid(self.env, vehicle=self.observer_vehicle,
time_quantization=1/self.env.config["policy_frequency"], horizon=self.horizon)
padding = np.ones(np.shape(grid))
padded_grid = np.concatenate([padding, grid, padding], axis=1)
obs_lanes = 3
l0 = grid.shape[1] + self.observer_vehicle.lane_index[2] - obs_lanes // 2
lf = grid.shape[1] + self.observer_vehicle.lane_index[2] + obs_lanes // 2
clamped_grid = padded_grid[:, l0:lf+1, :]
repeats = np.ones(clamped_grid.shape[0])
repeats[np.array([0, -1])] += clamped_grid.shape[0]
padded_grid = np.repeat(clamped_grid, repeats.astype(int), axis=0)
obs_speeds = 3
v0 = grid.shape[0] + self.observer_vehicle.speed_index - obs_speeds // 2
vf = grid.shape[0] + self.observer_vehicle.speed_index + obs_speeds // 2
clamped_grid = padded_grid[v0:vf + 1, :, :]
return np.append(obs.flatten(),clamped_grid.flatten(),axis=0)
def observe(self):
grid = compute_ttc_grid(self.env, time_quantization=1/self.env.config["policy_frequency"], horizon=self.horizon)
padding = np.ones(np.shape(grid))
padded_grid = np.concatenate([padding, grid, padding], axis=1)
obs_lanes = 3
l0 = grid.shape[1] + self.env.vehicle.lane_index[2] - obs_lanes // 2
lf = grid.shape[1] + self.env.vehicle.lane_index[2] + obs_lanes // 2
clamped_grid = padded_grid[:, l0:lf+1, :]
repeats = np.ones(clamped_grid.shape[0])
repeats[np.array([0, -1])] += clamped_grid.shape[0]
padded_grid = np.repeat(clamped_grid, repeats.astype(int), axis=0)
obs_velocities = 3
v0 = grid.shape[0] + self.env.vehicle.velocity_index - obs_velocities // 2
vf = grid.shape[0] + self.env.vehicle.velocity_index + obs_velocities // 2
clamped_grid = padded_grid[v0:vf + 1, :, :]
return clamped_grid
