def test_cost_function(plot=False):
# Create parameters
p = create_params()
obstacle_map = SBPDMap(p.obstacle_map_params)
obstacle_occupancy_grid = obstacle_map.create_occupancy_grid_for_map()
map_size_2 = obstacle_occupancy_grid.shape[::-1]
# Define a goal position and compute the corresponding fmm map
goal_pos_n2 = np.array([[20., 16.5]])
fmm_map = FmmMap.create_fmm_map_based_on_goal_position(goal_positions_n2=goal_pos_n2,
map_size_2=map_size_2,
dx=0.05,
map_origin_2=[0., 0.],
mask_grid_mn=obstacle_occupancy_grid)
# Define the cost function
objective_function = ObjectiveFunction(p.objective_fn_params)
objective_function.add_objective(ObstacleAvoidance(params=p.avoid_obstacle_objective, obstacle_map=obstacle_map))
objective_function.add_objective(GoalDistance(params=p.goal_distance_objective, fmm_map=fmm_map))
objective_function.add_objective(AngleDistance(params=p.goal_angle_objective, fmm_map=fmm_map))
# Define each objective separately
objective1 = ObstacleAvoidance(params=p.avoid_obstacle_objective, obstacle_map=obstacle_map)
objective2 = GoalDistance(params=p.goal_distance_objective, fmm_map=fmm_map)
objective3 = AngleDistance(params=p.goal_angle_objective, fmm_map=fmm_map)
# Define a set of positions and evaluate objective
pos_nk2 = tf.constant([[[8., 16.], [8., 12.5], [18., 16.5]]], dtype=tf.float32)
trajectory = Trajectory(dt=0.1, n=1, k=3, position_nk2=pos_nk2)
# Compute the objective function
values_by_objective = objective_function.evaluate_function_by_objective(trajectory)
overall_objective = objective_function.evaluate_function(trajectory)
# Expected objective values
expected_objective1 = objective1.evaluate_objective(trajectory)
expected_objective2 = objective2.evaluate_objective(trajectory)
expected_objective3 = objective3.evaluate_objective(trajectory)
expected_overall_objective = tf.reduce_mean(expected_objective1 + expected_objective2 + expected_objective3, axis=1)
# assert len(values_by_objective) == 3
# assert values_by_objective[0][1].shape == (1, 3)
# assert overall_objective.shape == (1,)
# assert np.allclose(overall_objective.numpy(), expected_overall_objective.numpy(), atol=1e-2)
# Optionally visualize the traversable and the points on which
# we compute the objective function
if plot:
import matplotlib.pyplot as plt
fig = plt.figure()
ax = fig.add_subplot(111)
obstacle_map.render(ax)
ax.plot(pos_nk2[0, :, 0].numpy(), pos_nk2[0, :, 1].numpy(), 'r.')
ax.plot(goal_pos_n2[0, 0], goal_pos_n2[0, 1], 'k*')
fig.savefig('./tmp/test_cost_function.png', bbox_inches='tight', pad_inches=0)
def test_avoid_obstacle():
# Create parameters
p = create_params()
# Create an SBPD Map
obstacle_map = SBPDMap(p.obstacle_map_params)
# Define the objective
objective = ObstacleAvoidance(params=p.avoid_obstacle_objective,
obstacle_map=obstacle_map)
# Define a set of positions and evaluate objective
pos_nk2 = tf.constant([[[8., 16.], [8., 12.5], [18., 16.5]]],
dtype=tf.float32)
trajectory = Trajectory(dt=0.1, n=1, k=3, position_nk2=pos_nk2)
# Compute the objective
objective_values_13 = objective.evaluate_objective(trajectory)
assert objective_values_13.shape == (1, 3)
# Expected objective values
distance_map = obstacle_map.fmm_map.fmm_distance_map.voxel_function_mn
idxs_xy_n2 = pos_nk2[0] / .05
idxs_yx_n2 = idxs_xy_n2[:, ::-1].numpy().astype(np.int32)
expected_min_dist_to_obs = np.array([
distance_map[idxs_yx_n2[0][0], idxs_yx_n2[0][1]],
distance_map[idxs_yx_n2[1][0], idxs_yx_n2[1][1]],
distance_map[idxs_yx_n2[2][0], idxs_yx_n2[2][1]]
],
dtype=np.float32)
m0 = p.avoid_obstacle_objective.obstacle_margin0
m1 = p.avoid_obstacle_objective.obstacle_margin1
expected_infringement = m1 - expected_min_dist_to_obs
expected_infringement = np.clip(expected_infringement, a_min=0,
a_max=None) # ReLU
expected_infringement /= (m1 - m0)
expected_objective = 25. * expected_infringement * expected_infringement
assert np.allclose(objective_values_13.numpy()[0],
expected_objective,
atol=1e-4)
assert np.allclose(objective_values_13.numpy()[0], [0., 0., 0.54201907],
atol=1e-4)
def _init_obj_fn(self, params=None):
"""
Initialize the objective function given sim params
"""
if params is None:
params = self.params
obstacle_map = self.obstacle_map
obj_fn = ObjectiveFunction(params.objective_fn_params)
if not params.avoid_obstacle_objective.empty():
obj_fn.add_objective(
ObstacleAvoidance(params=params.avoid_obstacle_objective,
obstacle_map=obstacle_map))
if not params.goal_distance_objective.empty():
obj_fn.add_objective(
GoalDistance(params=params.goal_distance_objective,
fmm_map=obstacle_map.fmm_map))
if not params.goal_angle_objective.empty():
obj_fn.add_objective(
AngleDistance(params=params.goal_angle_objective,
fmm_map=obstacle_map.fmm_map))
return obj_fn
def _init_obj_fn(self):
"""
Initialize the objective function.
Use fmm_map = None as this is undefined
until a goal configuration is specified.
"""
p = self.params
obj_fn = ObjectiveFunction(p.objective_fn_params)
if not p.avoid_obstacle_objective.empty():
obj_fn.add_objective(ObstacleAvoidance(
params=p.avoid_obstacle_objective,
obstacle_map=self.obstacle_map))
if not p.goal_distance_objective.empty():
obj_fn.add_objective(GoalDistance(
params=p.goal_distance_objective,
fmm_map=None))
if not p.goal_angle_objective.empty():
obj_fn.add_objective(AngleDistance(
params=p.goal_angle_objective,
fmm_map=None))
return obj_fn
def _init_obj_fn(self):
"""
Initialize the objective function.
Use fmm_map = None as this is undefined
until a goal configuration is specified.
"""
p = self.params
obj_fn = ObjectiveFunction(p.objective_fn_params)
if self.params.cost == 'heuristics':
if not p.avoid_obstacle_objective.empty():
obj_fn.add_objective(ObstacleAvoidance(
params=p.avoid_obstacle_objective,
obstacle_map=self.obstacle_map))
if not p.goal_angle_objective.empty():
obj_fn.add_objective(AngleDistance(
params=p.goal_angle_objective,
fmm_map=None))
if not p.goal_distance_objective.empty():
obj_fn.add_objective(GoalDistance(
params=p.goal_distance_objective,
fmm_map=None))
elif self.params.cost == 'reachability':
if not p.avoid_4d_objective.empty():
obj_fn.add_objective(Avoid4d(
params=p.avoid_4d_objective,
reachability_map=None))
if not p.goal_distance_objective.empty():
obj_fn.add_objective(GoalDistance(
params=p.goal_distance_objective,
fmm_map=None))
if not p.reach_avoid_4d_objective.empty():
obj_fn.add_objective(ReachAvoid4d(
params=p.reach_avoid_4d_objective,
reachability_map=None))
return obj_fn
def test_avoid_obstacle(visualize=False):
"""
Run a test on handpicked waypoint values that build a trajectory and are used
to evaluate the computed distance_map and angle_map from the fmm_*_map.voxel_function_mn
"""
# Create parameters
p = create_params()
# Create an SBPD Map
r, dx_cm, traversible = load_building(p)
obstacle_map = SBPDMap(p.obstacle_map_params,
renderer=0,
res=dx_cm,
map_trav=traversible)
# Define the objective
objective = ObstacleAvoidance(params=p.avoid_obstacle_objective,
obstacle_map=obstacle_map)
# Define a set of positions and evaluate objective
pos_nk2 = p.pos_nk2
trajectory = Trajectory(dt=0.1, n=1, k=3, position_nk2=pos_nk2)
# Compute the objective
objective_values_13 = objective.evaluate_objective(trajectory)
assert objective_values_13.shape == (1, 3)
# Expected objective values
distance_map = obstacle_map.fmm_map.fmm_distance_map.voxel_function_mn
angle_map = obstacle_map.fmm_map.fmm_angle_map.voxel_function_mn
idxs_xy_n2 = pos_nk2[0] / .05
idxs_yx_n2 = idxs_xy_n2[:, ::-1].astype(np.int32)
expected_min_dist_to_obs = np.array([
distance_map[idxs_yx_n2[0][0], idxs_yx_n2[0][1]],
distance_map[idxs_yx_n2[1][0], idxs_yx_n2[1][1]],
distance_map[idxs_yx_n2[2][0], idxs_yx_n2[2][1]]
],
dtype=np.float32)
m0 = p.avoid_obstacle_objective.obstacle_margin0
m1 = p.avoid_obstacle_objective.obstacle_margin1
expected_infringement = m1 - expected_min_dist_to_obs
expected_infringement = np.clip(expected_infringement, a_min=0,
a_max=None) # ReLU
expected_infringement /= (m1 - m0)
expected_objective = 25. * expected_infringement * expected_infringement
assert np.allclose(objective_values_13[0], expected_objective, atol=1e-4)
assert np.allclose(objective_values_13[0], p.test_obs_obj_ans, atol=1e-4)
if visualize:
"""
create a 1 x 3 (or 1 x 4) image of the obstacle map itself (as a traversible plot),
next to its corresponding angle_map, and distance_map. Optionally plotting the trajectory
"""
fig = plt.figure()
ax = fig.add_subplot(1, 3, 1)
obstacle_map.render(ax)
ax.plot(pos_nk2[0, :, 0], pos_nk2[0, :, 1], 'r.')
# ax.plot(objective[0, 0], objective[0, 1], 'k*')
ax.set_title('obstacle map')
# Plotting the "angle map"
ax = fig.add_subplot(1, 3, 2)
ax.imshow(angle_map, origin='lower')
ax.set_title('angle map')
# Plotting the "distance map"
ax = fig.add_subplot(1, 3, 3)
ax.imshow(distance_map, origin='lower')
ax.set_title('distance map')
# Plotting the trajectory
#ax = fig.add_subplot(1,4,4)
# trajectory.render(ax)
# ax.set_title('trajectory')
fig.savefig('./tests/obstacles/test_obstacle_objective.png',
bbox_inches='tight',
pad_inches=0)
def test_cost_function(plot=False):
"""
Creating objective points maually, plotting them in the ObjectiveFunction
class, and then asserting that combined, their sum adds up to the same
objective cost as the sum of the individual trajectories
"""
# Create parameters
p = create_params()
r, dx_cm, traversible = load_building(p)
obstacle_map = SBPDMap(p.obstacle_map_params,
renderer=0,
res=dx_cm,
map_trav=traversible)
# obstacle_map = SBPDMap(p.obstacle_map_params)
obstacle_occupancy_grid = obstacle_map.create_occupancy_grid_for_map()
map_size_2 = obstacle_occupancy_grid.shape[::-1]
# Define a goal position and compute the corresponding fmm map
goal_pos_n2 = p.goal_pos_n2
fmm_map = FmmMap.create_fmm_map_based_on_goal_position(
goal_positions_n2=goal_pos_n2,
map_size_2=map_size_2,
dx=0.05,
map_origin_2=[0., 0.],
mask_grid_mn=obstacle_occupancy_grid)
# Define the cost function
objective_function = ObjectiveFunction(p.objective_fn_params)
objective_function.add_objective(
ObstacleAvoidance(params=p.avoid_obstacle_objective,
obstacle_map=obstacle_map))
objective_function.add_objective(
GoalDistance(params=p.goal_distance_objective, fmm_map=fmm_map))
objective_function.add_objective(
AngleDistance(params=p.goal_angle_objective, fmm_map=fmm_map))
# Define each objective separately
objective1 = ObstacleAvoidance(params=p.avoid_obstacle_objective,
obstacle_map=obstacle_map)
objective2 = GoalDistance(params=p.goal_distance_objective,
fmm_map=fmm_map)
objective3 = AngleDistance(params=p.goal_angle_objective, fmm_map=fmm_map)
# Define a set of positions and evaluate objective
pos_nk2 = p.pos_nk2
heading_nk2 = np.array([[[np.pi / 2.0], [0.1], [0.1]]], dtype=np.float32)
trajectory = Trajectory(dt=0.1,
n=1,
k=3,
position_nk2=pos_nk2,
heading_nk1=heading_nk2)
# Compute the objective function
values_by_objective = objective_function.evaluate_function_by_objective(
trajectory)
overall_objective = objective_function.evaluate_function(trajectory)
# Expected objective values
expected_objective1 = objective1.evaluate_objective(trajectory)
expected_objective2 = objective2.evaluate_objective(trajectory)
expected_objective3 = objective3.evaluate_objective(trajectory)
# expected_overall_objective = tf.reduce_mean(
# expected_objective1 + expected_objective2 + expected_objective3, axis=1)
expected_overall_objective = np.mean(
expected_objective1 + expected_objective2 + expected_objective3,
axis=1)
assert len(values_by_objective) == 3
assert values_by_objective[0][1].shape == (1, 3)
assert overall_objective.shape == (1, )
# assert np.allclose(overall_objective.numpy(), expected_overall_objective.numpy(), atol=1e-2)
assert np.allclose(overall_objective,
expected_overall_objective,
atol=1e-2)
# Optionally visualize the traversable and the points on which
# we compute the objective function
if plot:
fig = plt.figure()
ax = fig.add_subplot(1, 1, 1)
obstacle_map.render(ax)
ax.plot(pos_nk2[0, :, 0].numpy(), pos_nk2[0, :, 1].numpy(), 'r.')
ax.plot(goal_pos_n2[0, 0], goal_pos_n2[0, 1], 'k*')
fig.savefig('./tests/cost/test_cost_function.png',
bbox_inches='tight',
pad_inches=0)
def test_personal_cost_function(sim_state: SimState,
plot=False,
verbose=False):
"""
Creating objective points maually, plotting them in the ObjectiveFunction
class, and then asserting that combined, their sum adds up to the same
objective cost as the sum of the individual trajectories
"""
# Create parameters
p = create_params()
r, dx_cm, traversible = load_building(p)
obstacle_map = SBPDMap(p.obstacle_map_params,
renderer=0,
res=dx_cm,
map_trav=traversible)
# obstacle_map = SBPDMap(p.obstacle_map_params)
obstacle_occupancy_grid = obstacle_map.create_occupancy_grid_for_map()
map_size_2 = obstacle_occupancy_grid.shape[::-1]
# Define a goal position and compute the corresponding fmm map
goal_pos_n2 = np.array([[9., 10]])
fmm_map = FmmMap.create_fmm_map_based_on_goal_position(
goal_positions_n2=goal_pos_n2,
map_size_2=map_size_2,
dx=0.05,
map_origin_2=[0., 0.],
mask_grid_mn=obstacle_occupancy_grid)
# Define the cost function
objective_function = ObjectiveFunction(p.objective_fn_params)
objective_function.add_objective(
ObstacleAvoidance(params=p.avoid_obstacle_objective,
obstacle_map=obstacle_map))
objective_function.add_objective(
GoalDistance(params=p.goal_distance_objective, fmm_map=fmm_map))
objective_function.add_objective(
AngleDistance(params=p.goal_angle_objective, fmm_map=fmm_map))
# Define each objective separately
objective1 = ObstacleAvoidance(params=p.avoid_obstacle_objective,
obstacle_map=obstacle_map)
objective2 = GoalDistance(params=p.goal_distance_objective,
fmm_map=fmm_map)
objective3 = AngleDistance(params=p.goal_angle_objective, fmm_map=fmm_map)
# Define cost function for personal state
objectiveP = PersonalSpaceCost(params=p.personal_space_params)
# Define a set of positions and evaluate objective
pos_nk2 = np.array([[[8., 12.5], [8., 16.], [18., 16.5]]],
dtype=np.float32)
heading_nk2 = np.array([[[np.pi / 2.0], [0.1], [0.1]]], dtype=np.float32)
trajectory = Trajectory(dt=0.1,
n=1,
k=3,
position_nk2=pos_nk2,
heading_nk1=heading_nk2)
# Compute the objective function
values_by_objective = objective_function.evaluate_function_by_objective(
trajectory)
overall_objective = objective_function.evaluate_function(trajectory)
# Expected objective values
expected_objective1 = objective1.evaluate_objective(trajectory)
expected_objective2 = objective2.evaluate_objective(trajectory)
expected_objective3 = objective3.evaluate_objective(trajectory)
expected_overall_objective = np.mean(
expected_objective1 + expected_objective2 + expected_objective3,
axis=1)
assert len(values_by_objective) == 3
assert values_by_objective[0][1].shape == (1, 3)
assert overall_objective.shape == (1, )
# assert np.allclose(overall_objective.numpy(), expected_overall_objective.numpy(), atol=1e-2)
assert np.allclose(overall_objective,
expected_overall_objective,
atol=1e-2)
# Use sim_state from main
sim_state_hist = {}
sim_state_hist[0] = sim_state
ps_cost = objectiveP.evaluate_objective(trajectory, sim_state_hist)
if verbose:
print("Personal space cost ", ps_cost)
print("Obstacle avoidance cost", expected_objective1)
print("Goal distance cost", expected_objective2)
print("Angle distance cost", expected_objective3)
# Optionally visualize the traversable and the points on which
# we compute the objective function
if plot:
fig = plt.figure()
ax = fig.add_subplot(1, 1, 1)
obstacle_map.render(ax)
# plot agent start
ax.plot(pos_nk2[0, :, 0], pos_nk2[0, :, 1], 'r.')
# plot agent goal
ax.plot(goal_pos_n2[0, 0], goal_pos_n2[0, 1], 'k*')
agents = sim_state.get_all_agents()
for agent_name, agent_vals in agents.items():
agent_pos3 = agent_vals.get_pos3() # (x,y,th)
theta = agent_pos3[2]
ax.plot(agent_pos3[0], agent_pos3[1], 'g.')
# plot non ego agents
fig.savefig('../test_psc_function.png',
bbox_inches='tight',
pad_inches=0)