RoboRally/remote_control/casadi_opt.py

317 lines
11 KiB
Python

from casadi import *
import time
# look at: https://github.com/casadi/casadi/blob/master/docs/examples/python/vdp_indirect_multiple_shooting.py
# TODO for roborally
# implement essential movements: forward, backward, turn left, turn right on grid-based level
class OpenLoopSolver:
def __init__(self, N=20, T=4.0):
self.T = T
self.N = N
self.opti_x0 = None
self.opti_lam_g0 = None
self.use_warmstart = True
def setup(self):
x = SX.sym('x')
y = SX.sym('y')
theta = SX.sym('theta')
state = vertcat(x, y, theta)
r = 0.03
R = 0.05
d = 0.02
#omega_max = 13.32
omega_max = 10.0
omegar = SX.sym('omegar')
omegal = SX.sym('omegal')
control = vertcat(omegar, omegal)
# model equation
f1 = (r / 2 * cos(theta) - r * d / (2 * R) * sin(theta)) * omegar * omega_max + (r / 2 * cos(theta) + r * d / (2 * R) * sin(
theta)) * omegal * omega_max
f2 = (r / 2 * sin(theta) + r * d / (2 * R) * cos(theta)) * omegar * omega_max + (r / 2 * sin(theta) - r * d / (2 * R) * cos(
theta)) * omegal * omega_max
f3 = -(r / (2 * R) * omegar - r / (2 * R) * omegal) * omega_max
xdot = vertcat(f1, f2, f3)
f = Function('f', [x, y, theta, omegar, omegal], [f1, f2, f3])
print("f = {}".format(f))
# cost functional
target = (-0.0, 0.0)
L = (x-target[0]) ** 2 + (y-target[1]) ** 2 + 1e-2 * (omegar ** 2 + omegal ** 2)
#L = (x-target[0]) ** 2 + (y-target[1]) ** 2 + 1e-2 * theta ** 2 + 1e-2 * (omegar ** 2 + omegal ** 2)
# Fixed step Runge-Kutta 4 integrator
M = 4 # RK4 steps per interval
DT = self.T / self.N / M
print("DT = {}".format(DT))
self.f = Function('f', [state, control], [xdot, L])
X0 = MX.sym('X0', 3)
U = MX.sym('U', 2)
X = X0
Q = 0
runge_kutta = True
if runge_kutta:
for j in range(M):
k1, k1_q = self.f(X, U)
k2, k2_q = self.f(X + DT / 2 * k1, U)
k3, k3_q = self.f(X + DT / 2 * k2, U)
k4, k4_q = self.f(X + DT * k3, U)
X = X + DT / 6 * (k1 + 2 * k2 + 2 * k3 + k4)
Q = Q + DT / 6 * (k1_q + 2 * k2_q + 2 * k3_q + k4_q)
else:
DT = self.T / self.N
k1, k1_q = f(X, U)
X = X + DT * k1
Q = Q + DT * k1_q
F = Function('F', [X0, U], [X, Q], ['x0', 'p'], ['xf', 'qf'])
# F_euler = Function('F_euler', [X0, U], [Xeuler, Qeuler], ['x0', 'p'], ['xf', 'qf'])
Fk = F(x0=[0.2, 0.3, 0.0], p=[-1.1, 1.1])
print(Fk['xf'])
print(Fk['qf'])
# Start with an empty NLP
w = []
w0 = []
lbw = []
ubw = []
J = 0
g = []
lbg = []
ubg = []
# Formulate the NLP
# Xk = MX(x0)
# for k in range(self.N):
# # New NLP variable for the control
# U1k = MX.sym('U1_' + str(k), 2)
# # U2k = MX.sym('U2_' + str(k))
# w += [U1k]
# lbw += [-0.5, -0.5]
# ubw += [0.5, 0.5]
# w0 += [0, 0]
#
# # Integrate till the end of the interval
# Fk = F(x0=Xk, p=U1k)
# Xk = Fk['xf']
# J = J + Fk['qf']
#
# # Add inequality constraint
# # g += [Xk[1]]
# # lbg += [-.0]
# # ubg += [inf]
#
# # Create an NLP solver
# prob = {'f': J, 'x': vertcat(*w), 'g': vertcat(*g)}
# self.solver = nlpsol('solver', 'ipopt', prob)
# Solve the NLP
if False:
sol = self.solver(x0=w0, lbx=lbw, ubx=ubw, lbg=lbg, ubg=ubg)
w_opt = sol['x']
# Plot the solution
u_opt = w_opt
x_opt = [self.x0]
for k in range(self.N):
Fk = F(x0=x_opt[-1], p=u_opt[2*k:2*k+2])
x_opt += [Fk['xf'].full()]
x1_opt = [r[0] for r in x_opt]
x2_opt = [r[1] for r in x_opt]
x3_opt = [r[2] for r in x_opt]
tgrid = [self.T/self.N*k for k in range(self.N+1)]
#import matplotlib.pyplot as plt
#plt.figure(2)
#plt.clf()
#plt.plot(tgrid, x1_opt, '--')
#plt.plot(tgrid, x2_opt, '-')
#plt.plot(tgrid, x3_opt, '*')
#plt.step(tgrid, vertcat(DM.nan(1), u_opt), '-.')
#plt.xlabel('t')
#plt.legend(['x1','x2','x3','u'])
#plt.grid()
#plt.show()
#return
def solve(self, x0, target):
angles_unwrapped = np.unwrap([x0[2], target[2]]) # unwrap angle to avoid jump in data
x0[2] = angles_unwrapped[0]
target[2] = angles_unwrapped[1]
tstart = time.time()
# alternative solution using multiple shooting (way faster!)
self.opti = Opti() # Optimization problem
# ---- decision variables ---------
self.X = self.opti.variable(3,self.N+1) # state trajectory
self.Q = self.opti.variable(1,self.N+1) # state trajectory
self.U = self.opti.variable(2,self.N) # control trajectory (throttle)
self.slack = self.opti.variable(1,1)
#T = self.opti.variable() # final time
# ---- objective ---------
#self.opti.minimize(T) # race in minimal time
# ---- dynamic constraints --------
#f = lambda x,u: vertcat(f1, f2, f3) # dx/dt = f(x,u)
# ---- initial values for solver ---
#self.opti.set_initial(speed, 1)
#self.opti.set_initial(T, 1)
x = SX.sym('x')
y = SX.sym('y')
theta = SX.sym('theta')
state = vertcat(x, y, theta)
r = 0.03
R = 0.05
d = 0.02
omega_max = 13.32
omegar = SX.sym('omegar')
omegal = SX.sym('omegal')
control = vertcat(omegar, omegal)
# model equation
f1 = (r / 2 * cos(theta) - r * d / (2 * R) * sin(theta)) * omegar * omega_max + (r / 2 * cos(theta) + r * d / (
2 * R) * sin(
theta)) * omegal * omega_max
f2 = (r / 2 * sin(theta) + r * d / (2 * R) * cos(theta)) * omegar * omega_max + (r / 2 * sin(theta) - r * d / (
2 * R) * cos(
theta)) * omegal * omega_max
f3 = -(r / (2 * R) * omegar - r / (2 * R) * omegal) * omega_max
xdot = vertcat(f1, f2, f3)
#L = 100 * (x - target[0]) ** 2 + 100 * (y - target[1]) ** 2 + 1e-1 * (theta - target[2])**2 + 1e-2 * (omegar ** 2 + omegal ** 2)
L = 100 * (x - target[0]) ** 2 + 100 * (y - target[1]) ** 2 + 1e-1 * (target[2] - theta) ** 2 + 1e-2 * (
omegar ** 2 + omegal ** 2)
#L = (x - target[0]) ** 2 + (y - target[1]) ** 2 + 1e-2 * (theta - target[2]) ** 2 + 1e-2 * (omegar ** 2 + omegal ** 2)
self.f = Function('f', [state, control], [xdot, L])
# ---- solve NLP ------
# set numerical backend
# delete constraints
self.opti.subject_to()
# add new constraints
dt = self.T / self.N # length of a control interval
for k in range(self.N): # loop over control intervals
# Runge-Kutta 4 integration
k1, k1_q = self.f(self.X[:, k], self.U[:, k])
k2, k2_q = self.f(self.X[:, k] + dt / 2 * k1, self.U[:, k])
k3, k3_q = self.f(self.X[:, k] + dt / 2 * k2, self.U[:, k])
k4, k4_q = self.f(self.X[:, k] + dt * k3, self.U[:, k])
x_next = self.X[:, k] + dt / 6 * (k1 + 2 * k2 + 2 * k3 + k4)
q_next = self.Q[:, k] + dt / 6 * (k1_q + 2 * k2_q + 2 * k3_q + k4_q)
self.opti.subject_to(self.X[:, k + 1] == x_next) # close the gaps
self.opti.subject_to(self.Q[:, k + 1] == q_next) # close the gaps
self.opti.minimize(self.Q[:, self.N] + 1.0e5 * self.slack**2)
# ---- path constraints -----------
# limit = lambda pos: 1-sin(2*pi*pos)/2
# self.opti.subject_to(speed<=limit(pos)) # track speed limit
maxcontrol = 0.9
self.opti.subject_to(self.opti.bounded(-maxcontrol, self.U, maxcontrol)) # control is limited
# ---- boundary conditions --------
# self.opti.subject_to(speed[-1]==0) # .. with speed 0
self.opti.subject_to(self.Q[:, 0] == 0.0)
solver = self.opti.solver("ipopt", {'print_time': False}, {"print_level": 0})
# ---- misc. constraints ----------
# self.opti.subject_to(X[1,:]>=0) # Time must be positive
# self.opti.subject_to(X[2,:]<=4) # Time must be positive
# self.opti.subject_to(X[2,:]>=-2) # Time must be positive
posx = self.X[0, :]
posy = self.X[1, :]
angle = self.X[2, :]
self.opti.subject_to(posx[0] == x0[0]) # start at position 0 ...
self.opti.subject_to(posy[0] == x0[1]) # ... from stand-still
self.opti.subject_to(angle[0] == x0[2]) # finish line at position 1
tend = time.time()
#print("setting up problem took {} seconds".format(tend - tstart))
tstart = time.time()
if self.use_warmstart and self.opti_x0 is not None:
try:
self.opti.set_initial(self.opti.lam_g, self.opti_lam_g0)
self.opti.set_initial(self.opti.x, self.opti_x0)
except RuntimeError:
print("could not set warmstart")
sol = self.opti.solve() # actual solve
tend = time.time()
#print("solving the problem took {} seconds".format(tend - tstart))
tstart = time.time()
self.opti_x0 = sol.value(self.opti.x)
self.opti_lam_g0 = sol.value(self.opti.lam_g)
#u_opt_1 = map(lambda x: float(x), [u_opt[i * 2] for i in range(0, 60)])
#u_opt_2 = map(lambda x: float(x), [u_opt[i * 2 + 1] for i in range(0, 60)])
u_opt_1 = sol.value(self.U[0,:])
u_opt_2 = sol.value(self.U[1,:])
tend = time.time()
#print("postprocessing took {} seconds".format(tend - tstart))
return (u_opt_1, u_opt_2, sol.value(posx), sol.value(posy))
#lam_g0 = sol.value(self.opti.lam_g)
#self.opti.solve()
from pylab import plot, step, figure, legend, show, spy
plot(sol.value(posx),label="posx")
plot(sol.value(posy),label="posy")
plot(sol.value(angle),label="angle")
plt.figure(3)
plot(sol.value(posx), sol.value(posy))
ax = plt.gca()
#circle = plt.Circle(p, r)
#ax.add_artist(circle)
#plot(limit(sol.value(pos)),'r--',label="speed limit")
#step(range(N),sol.value(U),'k',label="throttle")
legend(loc="upper left")
plt.show()
pass
# linearization
# A = zeros(states.shape[0], states.shape[0])
# for i in range(f.shape[0]):
# for j in range(states.shape[0]):
# A[i,j] = diff(f[i,0], states[j])
# Alin = A.subs([(theta,0), (omegar,0), (omegal,0)])
# print("A = {}".format(Alin))
# B = zeros(f.shape[0], controls.shape[0])
# for i in range(f.shape[0]):
# for j in range(controls.shape[0]):
# B[i,j] = diff(f[i,0], controls[j])
# print("B = {}".format(B))
# dfdtheta = diff(f, theta)
#print(dfdtheta.doit())
# takeaway: linearization is not helpful, because the linearized system is not stabilizable
# -> alternative: use nonlinear control method