RoboRally/remote_control/pid_controller.py

293 lines
12 KiB
Python
Raw Normal View History

import numpy as np
import math
import time
class PIDController:
def __init__(self, estimator):
self.t = time.time()
self.estimator = estimator
self.controlling = False
self.P_angle = 0.5 # PID parameters determined by Ziegler-Nichols. measured K_crit = 1.4, T_crit = 1.5
self.I_angle = 0.4
self.D_angle = 0.1
self.P_pos = 0.50
self.I_pos = 0.3
self.D_pos = 0.1
self.mode = None
def move_to_pos(self, target_pos, robot, near_target_counter=5):
near_target = 0
while near_target < near_target_counter:
while not self.estimator.event_queue.empty():
event = self.estimator.event_queue.get()
print("event: ", event)
if event[0] == 'click':
pass
elif event[0] == 'key':
key = event[1]
if key == 84: # arrow up
self.controlling = True
self.t = time.time()
elif key == 82: # arrow down
self.controlling = False
robot.send_cmd()
elif key == 48: # 0
target_pos = np.array([0.0, 0.0, 0.0])
elif key == 43: # +
self.control_scaling += 0.1
self.control_scaling = min(self.control_scaling, 1.0)
print("control scaling = ", self.control_scaling)
elif key == 45: # -
self.control_scaling -= 0.1
self.control_scaling = max(self.control_scaling, 0.1)
print("control scaling = ", self.control_scaling)
elif key == 113:
print("quit!")
self.controlling = False
robot.send_cmd()
return
elif key == 27: # escape
print("quit!")
self.controlling = False
robot.send_cmd()
return
x_pred = self.get_measurement(robot.id)
if x_pred is not None:
error_pos = np.linalg.norm(x_pred[0:2] - target_pos[0:2])
angles_unwrapped = np.unwrap([x_pred[2], target_pos[2]]) # unwrap angle to avoid jump in data
error_ang = np.abs(angles_unwrapped[0] - angles_unwrapped[1])
# print("error pos = ", error_pos)
# print("error_pos = {}, error_ang = {}".format(error_pos, error_ang))
# if error_pos > 0.075 or error_ang > 0.35:
if error_pos > 0.05 or error_ang > 0.1:
# solve mpc open loop problem
res = self.ols.solve(x_pred, target_pos)
# us1 = res[0]
# us2 = res[1]
us1 = res[0] * self.control_scaling
us2 = res[1] * self.control_scaling
# print("u = {}", (us1, us2))
# print("---------------- mpc solution took {} seconds".format(tmpc_end - tmpc_start))
dt_mpc = time.time() - self.t
if dt_mpc < self.dt: # wait until next control can be applied
# print("sleeping for {} seconds...".format(self.dt - dt_mpc))
time.sleep(self.dt - dt_mpc)
else:
us1 = [0] * self.mstep
us2 = [0] * self.mstep
near_target += 1
# send controls to the robot
for i in range(0, self.mstep): # option to use multistep mpc if len(range) > 1
u1 = us1[i]
u2 = us2[i]
robot.send_cmd(u1, u2)
if i < self.mstep:
time.sleep(self.dt)
self.t = time.time() # save time the most recent control was applied
else:
print("robot not detected yet!")
def interactive_control(self, robots):
controlled_robot_number = 0
robot = robots[controlled_robot_number]
ts = []
angles = []
target_pos = np.array([0.0, 0.0, 0.0])
e_angle_old = 0.0
e_pos_old = 0.0
i = 0.0
i_angle = 0.0
i_pos = 0.0
t0 = time.time()
running = True
while running:
# handle events from opencv window
while not self.estimator.event_queue.empty():
event = self.estimator.event_queue.get()
print("event: ", event)
if event[0] == 'click':
target_pos = event[1]
i_angle = 0
i_pos = 0
e_pos_old = 0
e_angle_old = 0
self.mode = 'combined'
elif event[0] == 'key':
key = event[1]
if key == 32: # arrow up
self.controlling = not self.controlling
if not self.controlling:
print("disable control")
robot.send_cmd() # stop robot
self.mode = None # reset control mode
else:
print("enable control")
i = 0
self.t = time.time()
elif key == 48: # 0
target_pos = np.array([0.0, 0.0, 0.0]) # TODO: use center of board for target pos
elif key == 97: # a
self.mode = 'angle'
e_angle_old = 0
i = 0
self.t = time.time()
elif key == 99: # c
self.mode = 'combined'
e_angle_old = 0
e_pos_old = 0
i_angle = 0
i_pos = 0
self.t = time.time()
elif key == 112: # p
self.mode = 'position'
e_pos_old = 0
i = 0
self.t = time.time()
elif key == 43: # +
self.P_pos += 0.1
print("P = ", self.P_angle)
elif key == 45: # -
self.P_pos -= 0.1
print("P = ", self.P_angle)
elif key == 9: # TAB
# switch controlled robot
robot.send_cmd() # stop current robot
controlled_robot_number = (controlled_robot_number + 1) % len(robots)
robot = robots[controlled_robot_number]
print(f"controlled robot: {robot.id}")
elif key == 113 or key == 27: # q or ESCAPE
print("quit!")
self.controlling = False
robot.send_cmd()
return
if self.controlling:
# measure state
x_pred = self.get_measurement(robot.id)
dt = self.t - time.time()
#print(f"x_pred = {x_pred}\ntarget = {target_pos}\nerror = {np.linalg.norm(target_pos - x_pred)}\n")
if self.mode == 'angle':
# compute angle such that robot faces to target point
target_angle = target_pos[2]
ts.append(time.time() - t0)
angles.append(x_pred[2])
angles_unwrapped = np.unwrap([x_pred[2], target_angle]) # unwrap angle to avoid jump in data
e_angle = angles_unwrapped[0] - angles_unwrapped[1] # angle difference
p = self.P_angle * e_angle
# i += self.I * dt * e # right Riemann sum
i += self.I_angle * dt * (e_angle + e_angle_old)/2.0 # trapezoidal rule
d = self.D_angle * (e_angle - e_angle_old)/dt
u1 = p - i - d
u2 = - u1
e_angle_old = e_angle
elif self.mode == 'position':
# we assume that we face straight to the target position
v = target_pos[0:2] - x_pred[0:2]
# check if robot faces in same direction as pos-target-vector
target_angle = math.atan2(v[1], v[0])
current_angle = x_pred[2]
angles_unwrapped = np.unwrap([current_angle, target_angle]) # unwrap angle to avoid jump in data
alpha = angles_unwrapped[0]
beta = angles_unwrapped[1]
epsilon = 0.3
if not (abs(alpha - beta) < epsilon or abs(alpha - beta + np.pi) < epsilon or abs(alpha - beta - np.pi) < epsilon):
print("switch to angle mode")
u1 = u2 = 0.0
robot.send_cmd(u1, u2)
self.mode = 'angle'
elif abs(alpha - beta) < epsilon:
# forward
e = np.linalg.norm(x_pred[0:2] - target_pos[0:2]) # compute euclidean distance to target
p = self.P * e
# i += self.I * dt * e # right Riemann sum
i += self.I * dt * (e + e_old) / 2.0 # trapezoidal rule
d = self.D * (e - e_old) / dt
u1 = p - i + d
u2 = u1
else:
# backward
e = np.linalg.norm(x_pred[0:2] - target_pos[0:2]) # compute euclidean distance to target
p = self.P * e
# i += self.I * dt * e # right Riemann sum
i += self.I * dt * (e + e_old) / 2.0 # trapezoidal rule
d = self.D * (e - e_old) / dt
u1 = -(p - i + d)
u2 = u1
e_old = e
#else:
# print("switching to angle mode")
# u1 = u2 = 0.0
# self.mode = 'angle'
elif self.mode == 'combined':
# compute angle such that robot faces to target point
v = target_pos[0:2] - x_pred[0:2]
target_angle = math.atan2(v[1], v[0])
angles_unwrapped = np.unwrap([x_pred[2], target_angle]) # unwrap angle to avoid jump in data
e_angle = angles_unwrapped[0] - angles_unwrapped[1] # angle difference
e_pos = np.linalg.norm(v)
p_angle = self.P_angle * e_angle
i_angle += self.I_angle * dt * (e_angle + e_angle_old) / 2.0 # trapezoidal rule
d_angle = self.D_angle * (e_angle - e_angle_old) / dt
p_pos = self.P_pos * e_pos
i_pos += self.I_pos * dt * (e_pos + e_pos_old) / 2.0 # trapezoidal rule
d_pos = self.D_pos * (e_pos - e_pos_old) / dt
u1 = p_angle + p_pos - i_angle - i_pos
u2 = - p_angle + p_pos + i_angle - i_pos
e_pos_old = e_pos
e_angle_old = e_angle
if e_pos < 0.05:
self.mode = 'angle'
else:
u1 = 0.0
u2 = 0.0
#print(f"u = ({u1}, {u2})")
robot.send_cmd(u1, u2)
self.t = time.time() # save time when the most recent control was applied
time.sleep(0.1)
def get_measurement(self, robot_id):
return self.estimator.get_robot_state_estimate(robot_id)