def expand_wpts(self):
n = 5 # number of pts per orig pt
dz = 1 / n
o_line = self.wpts
o_ss = self.ss
o_vs = self.vs
new_line = []
new_ss = []
new_vs = []
for i in range(self.N-1):
dd = lib.sub_locations(o_line[i+1], o_line[i])
for j in range(n):
pt = lib.add_locations(o_line[i], dd, dz*j)
new_line.append(pt)
ds = o_ss[i+1] - o_ss[i]
new_ss.append(o_ss[i] + dz*j*ds)
dv = o_vs[i+1] - o_vs[i]
new_vs.append(o_vs[i] + dv * j * dz)
self.wpts = np.array(new_line)
self.ss = np.array(new_ss)
self.vs = np.array(new_vs)
self.N = len(new_line)
def convert_pts_s_th(pts):
N = len(pts)
s_i = np.zeros(N - 1)
th_i = np.zeros(N - 1)
for i in range(N - 1):
s_i[i] = lib.get_distance(pts[i], pts[i + 1])
th_i[i] = lib.get_bearing(pts[i], pts[i + 1])
return s_i, th_i
def find_nvecs(self):
N = self.N
track = self.cline
nvecs = []
# new_track.append(track[0, :])
nvec = lib.theta_to_xy(np.pi / 2 +
lib.get_bearing(track[0, :], track[1, :]))
nvecs.append(nvec)
for i in range(1, len(track) - 1):
pt1 = track[i - 1]
pt2 = track[min((i, N)), :]
pt3 = track[min((i + 1, N - 1)), :]
th1 = lib.get_bearing(pt1, pt2)
th2 = lib.get_bearing(pt2, pt3)
if th1 == th2:
th = th1
else:
dth = lib.sub_angles_complex(th1, th2) / 2
th = lib.add_angles_complex(th2, dth)
new_th = th + np.pi / 2
nvec = lib.theta_to_xy(new_th)
nvecs.append(nvec)
nvec = lib.theta_to_xy(np.pi / 2 +
lib.get_bearing(track[-2, :], track[-1, :]))
nvecs.append(nvec)
self.nvecs = np.array(nvecs)
def print_update(self, plot_reward=True):
if self.ptr < 10:
return
mean10 = np.mean(self.rewards[self.ptr - 10:self.ptr])
mean100 = np.mean(self.rewards[max(0, self.ptr - 100):self.ptr])
# score = moving_average(self.rewards[self.ptr-100:self.ptr], 10)
print(
f"Run: {self.t_counter} --> Moving10: {mean10:.2f} --> Moving100: {mean100:.2f} "
)
if plot_reward:
lib.plot(self.rewards[0:self.ptr], figure_n=2)
def expand_centerline(self):
n = 2 # number of pts per orig pt
dz = 1 / n
o_line = self.cline
new_line = []
for i in range(self.N - 1):
dd = lib.sub_locations(o_line[i + 1], o_line[i])
for j in range(n):
pt = lib.add_locations(o_line[i], dd, dz * j)
new_line.append(pt)
self.cline = np.array(new_line)
self.N = len(new_line)
def load_center_pts(self):
track_data = []
filename = 'maps/' + self.map_name + '_std.csv'
try:
with open(filename, 'r') as csvfile:
csvFile = csv.reader(csvfile, quoting=csv.QUOTE_NONNUMERIC)
for lines in csvFile:
track_data.append(lines)
except FileNotFoundError:
raise FileNotFoundError("No map file center pts")
track = np.array(track_data)
print(f"Track Loaded: {filename} in reward")
N = len(track)
self.wpts = track[:, 0:2]
ss = np.array([lib.get_distance(self.wpts[i], self.wpts[i+1]) for i in range(N-1)])
ss = np.cumsum(ss)
self.ss = np.insert(ss, 0, 0)
self.total_s = self.ss[-1]
self.diffs = self.wpts[1:,:] - self.wpts[:-1,:]
self.l2s = self.diffs[:,0]**2 + self.diffs[:,1]**2
def test_kernel_pp():
conf = lib.load_config_namespace(config_file)
planner = PurePursuit(conf)
kernel = TrackKernel(conf)
safety_planner = Supervisor(planner, kernel, conf)
run_kernel_test(conf, safety_planner, 5, obstacles=0)
def __call__(self, s, a, s_p, dev):
r = collision_complete_reward(s_p)
pos_tt = np.array([s_p['poses_x'][0], s_p['poses_y'][0]])
pt_i, pt_ii, d_i, d_ii = find_closest_pt(pos_tt, self.wpts)
d = lib.get_distance(pt_i, pt_ii)
d_c = get_tiangle_h(d_i, d_ii, d) / self.dis_scale
th_ref = lib.get_bearing(pt_i, pt_ii)
th = s_p['poses_theta'][0]
d_th = abs(lib.sub_angles_complex(th_ref, th))
v_scale = s_p['linear_vels_x'][0] / self.max_v
r_race = self.mh * np.cos(d_th) * v_scale - self.md * d_c
return r + r_race
def test_rando_kernel():
conf = lib.load_config_namespace(config_file)
planner = RandomPlanner(conf)
kernel = TrackKernel(conf)
safety_planner = Supervisor(planner, kernel, conf)
run_kernel_test(conf, safety_planner, 5, obstacles=0)
def test_e2e():
agent_name = "EndToEnd_1"
conf = lib.load_config_namespace(config_file)
vehicle = TestEndToEnd(conf, agent_name)
run_multi_test(conf, vehicle)
def __call__(self, s, a, s_p, r) -> float:
if r == -1:
return r
else:
pt_i, pt_ii, d_i, d_ii = find_closest_pt(s_p[0:2], self.pts)
d = lib.get_distance(pt_i, pt_ii)
d_c = get_tiangle_h(d_i, d_ii, d) / self.dis_scale
th_ref = lib.get_bearing(pt_i, pt_ii)
th = s_p[2]
d_th = abs(lib.sub_angles_complex(th_ref, th))
v_scale = s_p[3] / self.max_v
shaped_r = distance_potential(s, s_p, self.end)
new_r = self.mh * np.cos(d_th) * v_scale - self.md * d_c
return new_r + r + shaped_r
def _expand_wpts(self):
n = 5 # number of pts per orig pt
#TODO: make this a parameter
dz = 1 / n
o_line = self.waypoints
o_vs = self.vs
new_line = []
new_vs = []
for i in range(len(o_line) - 1):
dd = lib.sub_locations(o_line[i + 1], o_line[i])
for j in range(n):
pt = lib.add_locations(o_line[i], dd, dz * j)
new_line.append(pt)
dv = o_vs[i + 1] - o_vs[i]
new_vs.append(o_vs[i] + dv * j * dz)
self.waypoints = np.array(new_line)
self.vs = np.array(new_vs)
def find_centerline(self, show=True):
dt = self.dt
d_search = 0.8
n_search = 11
dth = (np.pi * 4 / 5) / (n_search - 1)
# makes a list of search locations
search_list = []
for i in range(n_search):
th = -np.pi / 2 + dth * i
x = -np.sin(th) * d_search
y = np.cos(th) * d_search
loc = [x, y]
search_list.append(loc)
pt = start = np.array([self.conf.sx, self.conf.sy])
self.cline = [pt]
th = self.conf.stheta - np.pi / 2
while (lib.get_distance(pt, start) > d_search
or len(self.cline) < 10) and len(self.cline) < 500:
vals = []
self.search_space = []
for i in range(n_search):
d_loc = lib.transform_coords(search_list[i], -th)
search_loc = lib.add_locations(pt, d_loc)
self.search_space.append(search_loc)
x, y = self.xy_to_row_column(search_loc)
val = dt[y, x]
vals.append(val)
ind = np.argmax(vals)
d_loc = lib.transform_coords(search_list[ind], -th)
pt = lib.add_locations(pt, d_loc)
self.cline.append(pt)
if show:
self.plot_raceline_finding()
th = lib.get_bearing(self.cline[-2], pt)
print(f"Adding pt: {pt}")
self.cline = np.array(self.cline)
self.N = len(self.cline)
print(f"Raceline found")
if show:
self.plot_raceline_finding(True)
self.plot_raceline_finding(False)
def find_closest_pt(pt, wpts):
"""
Returns the two closes points in order along wpts
"""
dists = [lib.get_distance(pt, wpt) for wpt in wpts]
min_i = np.argmin(dists)
d_i = dists[min_i]
if min_i == len(dists) - 1:
min_i -= 1
if dists[max(min_i -1, 0) ] > dists[min_i+1]:
p_i = wpts[min_i]
p_ii = wpts[min_i+1]
d_i = dists[min_i]
d_ii = dists[min_i+1]
else:
p_i = wpts[min_i-1]
p_ii = wpts[min_i]
d_i = dists[min_i-1]
d_ii = dists[min_i]
return p_i, p_ii, d_i, d_ii
def test_fgm():
conf = lib.load_config_namespace(config_file)
vehicle = FollowTheGap(conf)
test_vehicle(conf, vehicle)
def train_e2e():
agent_name = "EndToEnd_1"
conf = lib.load_config_namespace(config_file)
vehicle = TrainEndToEnd(conf, agent_name, False)
TrainVehicle(conf, vehicle, 40000, 0)
def distance_potential(s, s_p, end, beta=0.2, scale=0.5):
prev_dist = lib.get_distance(s[0:2], end)
cur_dist = lib.get_distance(s_p[0:2], end)
d_dis = (prev_dist - cur_dist) / scale
return d_dis * beta
# plt.plot(t[:-1], f_lat)
# plt.plot(t[:-1], f_t, linewidth=3)
# plt.plot(t, np.ones_like(t) * f_max, '--')
# plt.plot(t, np.ones_like(t) * -f_max, '--')
# plt.plot(t, np.ones_like(t) * f_long_max, '--')
# plt.plot(t, np.ones_like(t) * -f_long_max, '--')
# plt.legend(['Flong', "f_lat", "f_t"])
return vs
def convert_pts_s_th(pts):
N = len(pts)
s_i = np.zeros(N - 1)
th_i = np.zeros(N - 1)
for i in range(N - 1):
s_i[i] = lib.get_distance(pts[i], pts[i + 1])
th_i[i] = lib.get_bearing(pts[i], pts[i + 1])
return s_i, th_i
if __name__ == "__main__":
# fname = "config_example_map"
fname = "config_test"
# fname = "vegas"
conf = lib.load_config_namespace(fname)
pre_map = PreMap(conf)
pre_map.run_conversion()
from GeneralTrainTest import TrainVehicle, run_multi_test
from auto_race_f110_gym.NavAgents.AgentSerial import SerialVehicleTrain, SerialVehicleTest
from auto_race_f110_gym.Utils import LibFunctions as lib
config_file = "config_test"
conf = lib.load_config_namespace(config_file)
n = 4
serial_name = f"SerialPlanner_Berlin_{n}"
def train_serial():
vehicle = SerialVehicleTrain(serial_name, conf, False)
# vehicle = SerialVehicleTrain(serial_name, conf, True)
TrainVehicle(conf, vehicle, 200000, 6)
def test_serial():
vehicle = SerialVehicleTest(serial_name, conf)
run_multi_test(conf, vehicle)
if __name__ == "__main__":
# train_serial()
test_serial()
def test_pure_pursuit():
conf = lib.load_config_namespace(config_file)
vehicle = PurePursuit(conf)
run_multi_test(conf, vehicle, 5, obstacles=0)
def MinCurvatureTrajectory(pts, nvecs, ws):
"""
This function uses optimisation to minimise the curvature of the path
"""
width_factor = 0.6
w_min = -ws[:, 0] * width_factor
w_max = ws[:, 1] * width_factor
th_ns = [lib.get_bearing([0, 0], nvecs[i, 0:2]) for i in range(len(nvecs))]
N = len(pts)
n_f_a = ca.MX.sym('n_f', N)
n_f = ca.MX.sym('n_f', N - 1)
th_f = ca.MX.sym('n_f', N - 1)
x0_f = ca.MX.sym('x0_f', N - 1)
x1_f = ca.MX.sym('x1_f', N - 1)
y0_f = ca.MX.sym('y0_f', N - 1)
y1_f = ca.MX.sym('y1_f', N - 1)
th1_f = ca.MX.sym('y1_f', N - 1)
th2_f = ca.MX.sym('y1_f', N - 1)
th1_f1 = ca.MX.sym('y1_f', N - 2)
th2_f1 = ca.MX.sym('y1_f', N - 2)
o_x_s = ca.Function('o_x', [n_f], [pts[:-1, 0] + nvecs[:-1, 0] * n_f])
o_y_s = ca.Function('o_y', [n_f], [pts[:-1, 1] + nvecs[:-1, 1] * n_f])
o_x_e = ca.Function('o_x', [n_f], [pts[1:, 0] + nvecs[1:, 0] * n_f])
o_y_e = ca.Function('o_y', [n_f], [pts[1:, 1] + nvecs[1:, 1] * n_f])
dis = ca.Function('dis', [x0_f, x1_f, y0_f, y1_f],
[ca.sqrt((x1_f - x0_f)**2 + (y1_f - y0_f)**2)])
track_length = ca.Function('length', [n_f_a], [
dis(o_x_s(n_f_a[:-1]), o_x_e(n_f_a[1:]), o_y_s(n_f_a[:-1]),
o_y_e(n_f_a[1:]))
])
real = ca.Function(
'real', [th1_f, th2_f],
[ca.cos(th1_f) * ca.cos(th2_f) + ca.sin(th1_f) * ca.sin(th2_f)])
im = ca.Function(
'im', [th1_f, th2_f],
[-ca.cos(th1_f) * ca.sin(th2_f) + ca.sin(th1_f) * ca.cos(th2_f)])
sub_cmplx = ca.Function('a_cpx', [th1_f, th2_f],
[ca.atan2(im(th1_f, th2_f), real(th1_f, th2_f))])
get_th_n = ca.Function(
'gth', [th_f],
[sub_cmplx(ca.pi * np.ones(N - 1), sub_cmplx(th_f, th_ns[:-1]))])
d_n = ca.Function('d_n', [n_f_a, th_f],
[track_length(n_f_a) / ca.tan(get_th_n(th_f))])
# objective
real1 = ca.Function(
'real1', [th1_f1, th2_f1],
[ca.cos(th1_f1) * ca.cos(th2_f1) + ca.sin(th1_f1) * ca.sin(th2_f1)])
im1 = ca.Function(
'im1', [th1_f1, th2_f1],
[-ca.cos(th1_f1) * ca.sin(th2_f1) + ca.sin(th1_f1) * ca.cos(th2_f1)])
sub_cmplx1 = ca.Function(
'a_cpx1', [th1_f1, th2_f1],
[ca.atan2(im1(th1_f1, th2_f1), real1(th1_f1, th2_f1))])
# define symbols
n = ca.MX.sym('n', N)
th = ca.MX.sym('th', N - 1)
nlp = {\
'x': ca.vertcat(n, th),
'f': ca.sumsqr(sub_cmplx1(th[1:], th[:-1])),
# 'f': ca.sumsqr(track_length(n)),
'g': ca.vertcat(
# dynamic constraints
n[1:] - (n[:-1] + d_n(n, th)),
# boundary constraints
n[0], #th[0],
n[-1], #th[-1],
) \
}
# S = ca.nlpsol('S', 'ipopt', nlp, {'ipopt':{'print_level':5}})
S = ca.nlpsol('S', 'ipopt', nlp, {'ipopt': {'print_level': 0}})
ones = np.ones(N)
n0 = ones * 0
th0 = []
for i in range(N - 1):
th_00 = lib.get_bearing(pts[i, 0:2], pts[i + 1, 0:2])
th0.append(th_00)
th0 = np.array(th0)
x0 = ca.vertcat(n0, th0)
lbx = list(w_min) + [-np.pi] * (N - 1)
ubx = list(w_max) + [np.pi] * (N - 1)
r = S(x0=x0, lbg=0, ubg=0, lbx=lbx, ubx=ubx)
x_opt = r['x']
n_set = np.array(x_opt[:N])
# thetas = np.array(x_opt[1*N:2*(N-1)])
return n_set
