-
Daniel Zhuang authoredDaniel Zhuang authored
mp0_p1.py 15.82 KiB
from typing import Tuple, List
import numpy as np
from scipy.integrate import ode
from verse import BaseAgent, Scenario
from verse.analysis.utils import wrap_to_pi
from verse.analysis.analysis_tree import TraceType, AnalysisTree
from verse.parser import ControllerIR
from vehicle_controller import VehicleMode, PedestrianMode
from verse.analysis import AnalysisTreeNode, AnalysisTree, AnalysisTreeNodeType
import copy
refine_profile = {
'R1': [0],
'R2': [0,0,0,3],
'R3': [0,0,0,3]
}
def tree_safe(tree: AnalysisTree):
for node in tree.nodes:
if node.assert_hits is not None:
return False
return True
def verify_refine(scenario: Scenario, time_horizon, time_step):
refine_depth = 5
init_car = scenario.init_dict['car']
init_ped = scenario.init_dict['pedestrian']
partition_depth = 0
if init_ped[1][0] - init_ped[0][0]>0.1:
exp = 'R3'
elif init_car[1][3] - init_car[0][3] > 0.1:
exp = 'R2'
else:
exp = 'R1'
res_list = []
init_queue = []
if False:
car_v_init_range = np.linspace(init_car[0][3], init_car[1][3], 5)
else:
car_v_init_range = [init_car[0][3], init_car[1][3]]
if init_car[1][0]-init_car[0][0] > 0.1:
# if False:
car_x_init_range = np.linspace(init_car[0][0], init_car[1][0], 5)
else:
car_x_init_range = [init_car[0][0], init_car[1][0]]
for i in range(len(car_x_init_range)-1):
for j in range(len(car_v_init_range)-1):
tmp = copy.deepcopy(init_car)
tmp[0][0] = car_x_init_range[i]
tmp[1][0] = car_x_init_range[i+1]
tmp[0][3] = car_v_init_range[j]
tmp[1][3] = car_v_init_range[j+1]
init_queue.append((tmp, init_ped, partition_depth))
# init_queue = [(init_car, init_ped, partition_depth)]
while init_queue!=[] and partition_depth < refine_depth:
car_init, ped_init, partition_depth = init_queue.pop(0)
print(f"######## {partition_depth}, car x, {car_init[0][0]}, {car_init[1][0]}, car v, {car_init[0][3]}, {car_init[1][3]}, ped x, {ped_init[0][0]}, {ped_init[1][0]}, ped y, {ped_init[0][1]}, {ped_init[1][1]}")
scenario.set_init_single('car', car_init, (VehicleMode.Normal,))
scenario.set_init_single('pedestrian', ped_init, (PedestrianMode.Normal,))
traces = scenario.verify(time_horizon, time_step, max_height=6)
if not tree_safe(traces):
# Partition car and pedestrian initial state
idx = refine_profile[exp][partition_depth%len(refine_profile[exp])]
if car_init[1][idx] - car_init[0][idx] < 0.01:
print(f"Stop refine car state {idx}")
init_queue.append((car_init, ped_init, partition_depth+1))
elif partition_depth >= refine_depth:
print('Threshold Reached. Scenario is UNSAFE.')
res_list.append(traces)
break
car_v_init = (car_init[0][idx] + car_init[1][idx])/2
car_init1 = copy.deepcopy(car_init)
car_init1[1][idx] = car_v_init
init_queue.append((car_init1, ped_init, partition_depth+1))
car_init2 = copy.deepcopy(car_init)
car_init2[0][idx] = car_v_init
init_queue.append((car_init2, ped_init, partition_depth+1))
else:
res_list.append(traces)
# com_traces = combine_tree(res_list)
return res_list
class PedestrianAgent(BaseAgent):
def __init__(
self,
id,
):
self.decision_logic: ControllerIR = ControllerIR.empty()
self.id = id
@staticmethod
def dynamic(t, state):
x, y, theta, v = state
x_dot = 0
y_dot = v
theta_dot = 0
v_dot = 0
return [x_dot, y_dot, theta_dot, v_dot]
def TC_simulate(
self, mode: List[str], init, time_bound, time_step, lane_map = None
) -> TraceType:
time_bound = float(time_bound)
num_points = int(np.ceil(time_bound / time_step))
trace = np.zeros((num_points + 1, 1 + len(init)))
trace[1:, 0] = [round(i * time_step, 10) for i in range(num_points)]
trace[0, 1:] = init
for i in range(num_points):
r = ode(self.dynamic)
r.set_initial_value(init)
res: np.ndarray = r.integrate(r.t + time_step)
init = res.flatten()
if init[3] < 0:
init[3] = 0
trace[i + 1, 0] = time_step * (i + 1)
trace[i + 1, 1:] = init
return trace
class VehicleAgent(BaseAgent):
def __init__(
self,
id,
code = None,
file_name = None,
accel_brake = 5,
accel_notbrake = 5,
accel_hardbrake = 20,
speed = 10
):
super().__init__(
id, code, file_name
)
self.accel_brake = accel_brake
self.accel_notbrake = accel_notbrake
self.accel_hardbrake = accel_hardbrake
self.speed = speed
self.vmax = 20
@staticmethod
def dynamic(t, state, u):
x, y, theta, v = state
delta, a = u
x_dot = v * np.cos(theta + delta)
y_dot = v * np.sin(theta + delta)
theta_dot = v / 1.75 * np.sin(delta)
v_dot = a
return [x_dot, y_dot, theta_dot, v_dot]
def action_handler(self, mode: List[str], state) -> Tuple[float, float]:
x, y, theta, v = state
vehicle_mode, = mode
vehicle_pos = np.array([x, y])
a = 0
lane_width = 3
d = -y
if vehicle_mode == "Normal" or vehicle_mode == "Stop":
pass
elif vehicle_mode == "SwitchLeft":
d += lane_width
elif vehicle_mode == "SwitchRight":
d -= lane_width
elif vehicle_mode == "Brake":
a = max(-self.accel_brake, -v)
# a = -50
elif vehicle_mode == "HardBrake":
a = max(-self.accel_hardbrake, -v)
# a = -50
elif vehicle_mode == "Accel":
a = min(self.accel_notbrake, self.speed-v)
else:
raise ValueError(f"Invalid mode: {vehicle_mode}")
heading = 0
psi = wrap_to_pi(heading - theta)
steering = psi + np.arctan2(0.45 * d, v)
steering = np.clip(steering, -0.61, 0.61)
return steering, a
def TC_simulate(
self, mode: List[str], init, time_bound, time_step, lane_map = None
) -> TraceType:
time_bound = float(time_bound)
num_points = int(np.ceil(time_bound / time_step))
trace = np.zeros((num_points + 1, 1 + len(init)))
trace[1:, 0] = [round(i * time_step, 10) for i in range(num_points)]
trace[0, 1:] = init
for i in range(num_points):
steering, a = self.action_handler(mode, init)
r = ode(self.dynamic)
r.set_initial_value(init).set_f_params([steering, a])
res: np.ndarray = r.integrate(r.t + time_step)
init = res.flatten()
if init[3] < 0:
init[3] = 0
trace[i + 1, 0] = time_step * (i + 1)
trace[i + 1, 1:] = init
return trace
def dist(pnt1, pnt2):
return np.linalg.norm(
np.array(pnt1) - np.array(pnt2)
)
def get_extreme(rect1, rect2):
lb11 = rect1[0]
lb12 = rect1[1]
ub11 = rect1[2]
ub12 = rect1[3]
lb21 = rect2[0]
lb22 = rect2[1]
ub21 = rect2[2]
ub22 = rect2[3]
# Using rect 2 as reference
left = lb21 > ub11
right = ub21 < lb11
bottom = lb22 > ub12
top = ub22 < lb12
if top and left:
dist_min = dist((ub11, lb12),(lb21, ub22))
dist_max = dist((lb11, ub12),(ub21, lb22))
elif bottom and left:
dist_min = dist((ub11, ub12),(lb21, lb22))
dist_max = dist((lb11, lb12),(ub21, ub22))
elif top and right:
dist_min = dist((lb11, lb12), (ub21, ub22))
dist_max = dist((ub11, ub12), (lb21, lb22))
elif bottom and right:
dist_min = dist((lb11, ub12),(ub21, lb22))
dist_max = dist((ub11, lb12),(lb21, ub22))
elif left:
dist_min = lb21 - ub11
dist_max = np.sqrt((lb21 - ub11)**2 + max((ub22-lb12)**2, (ub12-lb22)**2))
elif right:
dist_min = lb11 - ub21
dist_max = np.sqrt((lb21 - ub11)**2 + max((ub22-lb12)**2, (ub12-lb22)**2))
elif top:
dist_min = lb12 - ub22
dist_max = np.sqrt((ub12 - lb22)**2 + max((ub21-lb11)**2, (ub11-lb21)**2))
elif bottom:
dist_min = lb22 - ub12
dist_max = np.sqrt((ub22 - lb12)**2 + max((ub21-lb11)**2, (ub11-lb21)**2))
else:
dist_min = 0
dist_max = max(
dist((lb11, lb12), (ub21, ub22)),
dist((lb11, ub12), (ub21, lb22)),
dist((ub11, lb12), (lb21, ub12)),
dist((ub11, ub12), (lb21, lb22))
)
return dist_min, dist_max
class VehiclePedestrianSensor:
def __init__(self):
self.sensor_distance = 60
# The baseline sensor is omniscient. Each agent can get the state of all other agents
def sense(self, agent: BaseAgent, state_dict, lane_map):
len_dict = {}
cont = {}
disc = {}
len_dict = {"others": len(state_dict) - 1}
tmp = np.array(list(state_dict.values())[0][0])
if tmp.ndim < 2:
if agent.id == 'car':
len_dict['others'] = 1
cont['ego.x'] = state_dict['car'][0][1]
cont['ego.y'] = state_dict['car'][0][2]
cont['ego.theta'] = state_dict['car'][0][3]
cont['ego.v'] = state_dict['car'][0][4]
disc['ego.agent_mode'] = state_dict['car'][1][0]
dist = np.sqrt(
(state_dict['car'][0][1]-state_dict['pedestrian'][0][1])**2+\
(state_dict['car'][0][2]-state_dict['pedestrian'][0][2])**2
)
# cont['ego.dist'] = dist
if dist < self.sensor_distance:
cont['other.dist'] = dist
# cont['other.x'] = state_dict['pedestrian'][0][1]
# cont['other.y'] = state_dict['pedestrian'][0][2]
# cont['other.v'] = state_dict['pedestrian'][0][4]
else:
cont['other.dist'] = 1000
# cont['other.x'] = 1000
# cont['other.y'] = 1000
# cont['other.v'] = 1000
else:
if agent.id == 'car':
len_dict['others'] = 1
dist_min, dist_max = get_extreme(
(state_dict['car'][0][0][1],state_dict['car'][0][0][2],state_dict['car'][0][1][1],state_dict['car'][0][1][2]),
(state_dict['pedestrian'][0][0][1],state_dict['pedestrian'][0][0][2],state_dict['pedestrian'][0][1][1],state_dict['pedestrian'][0][1][2]),
)
cont['ego.x'] = [
state_dict['car'][0][0][1], state_dict['car'][0][1][1]
]
cont['ego.y'] = [
state_dict['car'][0][0][2], state_dict['car'][0][1][2]
]
cont['ego.theta'] = [
state_dict['car'][0][0][3], state_dict['car'][0][1][3]
]
cont['ego.v'] = [
state_dict['car'][0][0][4], state_dict['car'][0][1][4]
]
cont['other.dist'] = [
dist_min, dist_max
]
disc['ego.agent_mode'] = state_dict['car'][1][0]
if dist_min<self.sensor_distance:
cont['other.dist'] = [
dist_min, dist_max
]
# cont['other.x'] = [
# state_dict['pedestrian'][0][0][1], state_dict['pedestrian'][0][1][1]
# ]
# cont['other.y'] = [
# state_dict['pedestrian'][0][0][2], state_dict['pedestrian'][0][1][2]
# ]
# cont['other.v'] = [
# state_dict['pedestrian'][0][0][4], state_dict['pedestrian'][0][1][4]
# ]
else:
cont['other.dist'] = [
1000, 1000
]
# cont['other.x'] = [
# 1000, 1000
# ]
# cont['other.y'] = [
# 1000, 1000
# ]
# cont['other.v'] = [
# 1000, 1000
# ]
return cont, disc, len_dict
def sample_init(scenario: Scenario, num_sample=50):
"""
TODO: given the initial set,
generate multiple initial points located in the initial set
as the input of multiple simulation.
note that output should be formatted correctly and every point should be in inital set.
refer the following sample code to write your code.
"""
init_dict = scenario.init_dict
print(init_dict)
############## Your Code Start Here ##############
sample_dict_list = []
np.random.seed(2023)
for i in range(num_sample):
sample_dict={}
for agent in init_dict:
point = np.random.uniform(init_dict[agent][0], init_dict[agent][1]).tolist()
sample_dict[agent] = point
sample_dict_list.append(sample_dict)
############## Your Code End Here ##############
print(sample_dict_list)
return sample_dict_list
def eval_velocity(tree_list: List[AnalysisTree]):
agent_id = 'car'
velo_list = []
unsafe_init = []
for tree in tree_list:
assert agent_id in tree.root.init
leaves = list(filter(lambda node: node.child == [], tree.nodes))
unsafe = list(filter(lambda node: node.assert_hits != None, leaves))
if len(unsafe) != 0:
print(f"unsafety detected in tree with init {tree.root.init}")
unsafe_init.append(tree.root.init)
else:
safe = np.array(list(filter(lambda node: node.assert_hits == None, leaves)))
init_x = tree.root.init[agent_id][0]
last_xs = np.array([node.trace[agent_id][-1][1] for node in safe])
time = round(safe[0].trace[agent_id][-1][0], 3)
velos = (last_xs-init_x)/time
max_velo = np.max(velos)
velo_list.append(max_velo)
print(f"Max avg velocity {max_velo} in tree with init {tree.root.init}")
if len(tree_list) == len(velo_list):
print(f"No unsafety detected! Overall average velocity is {sum(velo_list)/len(velo_list)}.")
return {sum(velo_list)/len(velo_list)}, 0, []
else:
print(f"Unsafety detected! Please update your DL.")
return None, float(len(unsafe_init))/float(len(tree_list)), unsafe_init
def combine_tree(tree_list: List[AnalysisTree]):
combined_trace={}
for tree in tree_list:
for node in tree.nodes:
for agent_id in node.agent:
traces = node.trace
trace = np.array(traces[agent_id])
if agent_id not in combined_trace:
combined_trace[agent_id]={}
for i in range (0, len(trace), 2):
step = round(trace[i][0], 3)
if step not in combined_trace[agent_id]:
combined_trace[agent_id][step]=[trace[i], trace[i+1]]
else:
lower = np.min([combined_trace[agent_id][step][0],trace[i]], 0)
upper = np.max([combined_trace[agent_id][step][1],trace[i+1]], 0)
combined_trace[agent_id][step]=[lower, upper]
final_trace = {agent_id:np.array(list(combined_trace[agent_id].values())).flatten().reshape((-1, trace[i].size)).tolist() for agent_id in combined_trace}
root = AnalysisTreeNode(final_trace,None,tree_list[0].root.mode,None,None, node.agent, None,None,[],0,10,AnalysisTreeNodeType.REACH_TUBE,0)
return AnalysisTree(root)