Tuned superstructure loop and added feed forwards.
Change-Id: Ia2e3a1746529a4c27395f2e9b6e875c5cddb7616
diff --git a/y2016/control_loops/python/arm.py b/y2016/control_loops/python/arm.py
new file mode 100644
index 0000000..28a704d
--- /dev/null
+++ b/y2016/control_loops/python/arm.py
@@ -0,0 +1,407 @@
+#!/usr/bin/python
+
+import numpy
+import sys
+import operator
+
+from frc971.control_loops.python import control_loop
+from frc971.control_loops.python import controls
+
+from y2016.control_loops.python.shoulder import Shoulder, IntegralShoulder
+from y2016.control_loops.python.wrist import Wrist, IntegralWrist
+from aos.common.util.trapezoid_profile import TrapizoidProfile
+
+from matplotlib import pylab
+import gflags
+import glog
+
+FLAGS = gflags.FLAGS
+
+try:
+ gflags.DEFINE_bool('plot', False, 'If true, plot the loop response.')
+except gflags.DuplicateFlagError:
+ pass
+
+
+class Arm(control_loop.ControlLoop):
+ def __init__(self, name="Arm"):
+ super(Arm, self).__init__(name=name)
+ self._shoulder = Shoulder(name=name)
+ self._shooter = Wrist(name=name)
+
+ # Do a coordinate transformation.
+ # X_shooter_grounded = X_shooter + X_shoulder
+ # dX_shooter_grounded/dt = A_shooter * X_shooter + A_shoulder * X_shoulder +
+ # B_shoulder * U_shoulder + B_shooter * U_shooter
+ # dX_shooter_grounded/dt = A_shooter * (X_shooter_grounded - X_shoulder) +
+ # A_shoulder * X_shoulder + B_shooter * U_shooter + B_shoulder * U_shoulder
+ # X = [X_shoulder; X_shooter + X_shoulder]
+ # dX/dt = [A_shoulder 0] [X_shoulder ] + [B_shoulder 0] [U_shoulder]
+ # [(A_shoulder - A_shooter) A_shooter] [X_shooter_grounded] + [B_shoulder B_shooter] [ U_shooter]
+ # Y_shooter_grounded = Y_shooter + Y_shoulder
+
+ self.A_continuous = numpy.matrix(numpy.zeros((4, 4)))
+ self.A_continuous[0:2, 0:2] = self._shoulder.A_continuous
+ self.A_continuous[2:4, 0:2] = (self._shoulder.A_continuous -
+ self._shooter.A_continuous)
+ self.A_continuous[2:4, 2:4] = self._shooter.A_continuous
+
+ self.B_continuous = numpy.matrix(numpy.zeros((4, 2)))
+ self.B_continuous[0:2, 0:1] = self._shoulder.B_continuous
+ self.B_continuous[2:4, 1:2] = self._shooter.B_continuous
+ self.B_continuous[2:4, 0:1] = self._shoulder.B_continuous
+
+ self.C = numpy.matrix(numpy.zeros((2, 4)))
+ self.C[0:1, 0:2] = self._shoulder.C
+ self.C[1:2, 0:2] = -self._shoulder.C
+ self.C[1:2, 2:4] = self._shooter.C
+
+ # D is 0 for all our loops.
+ self.D = numpy.matrix(numpy.zeros((2, 2)))
+
+ self.dt = 0.005
+
+ self.A, self.B = self.ContinuousToDiscrete(
+ self.A_continuous, self.B_continuous, self.dt)
+
+ # Cost of error
+ self.Q = numpy.matrix(numpy.zeros((4, 4)))
+ q_pos_shoulder = 0.014
+ q_vel_shoulder = 4.00
+ q_pos_shooter = 0.014
+ q_vel_shooter = 4.00
+ self.Q[0, 0] = 1.0 / q_pos_shoulder ** 2.0
+ self.Q[1, 1] = 1.0 / q_vel_shoulder ** 2.0
+ self.Q[2, 2] = 1.0 / q_pos_shooter ** 2.0
+ self.Q[3, 3] = 1.0 / q_vel_shooter ** 2.0
+
+ # Cost of control effort
+ self.R = numpy.matrix(numpy.zeros((2, 2)))
+ r_voltage = 1.0 / 12.0
+ self.R[0, 0] = r_voltage ** 2.0
+ self.R[1, 1] = r_voltage ** 2.0
+
+ self.Kff = controls.TwoStateFeedForwards(self.B, self.Q)
+
+ glog.debug('Shoulder K')
+ glog.debug(self._shoulder.K)
+
+ # Compute controller gains.
+ # self.K = controls.dlqr(self.A, self.B, self.Q, self.R)
+ self.K = numpy.matrix(numpy.zeros((2, 4)))
+ self.K[0:1, 0:2] = self._shoulder.K
+ self.K[1:2, 0:2] = (
+ -self.Kff[1:2, 2:4] * self.B[2:4, 0:1] * self._shoulder.K
+ + self.Kff[1:2, 2:4] * self.A[2:4, 0:2])
+ self.K[1:2, 2:4] = self._shooter.K
+
+ glog.debug('Arm controller %s', repr(self.K))
+
+ # Cost of error
+ self.Q = numpy.matrix(numpy.zeros((4, 4)))
+ q_pos_shoulder = 0.05
+ q_vel_shoulder = 2.65
+ q_pos_shooter = 0.05
+ q_vel_shooter = 2.65
+ self.Q[0, 0] = q_pos_shoulder ** 2.0
+ self.Q[1, 1] = q_vel_shoulder ** 2.0
+ self.Q[2, 2] = q_pos_shooter ** 2.0
+ self.Q[3, 3] = q_vel_shooter ** 2.0
+
+ # Cost of control effort
+ self.R = numpy.matrix(numpy.zeros((2, 2)))
+ r_voltage = 0.025
+ self.R[0, 0] = r_voltage ** 2.0
+ self.R[1, 1] = r_voltage ** 2.0
+
+ self.KalmanGain, self.Q_steady = controls.kalman(
+ A=self.A, B=self.B, C=self.C, Q=self.Q, R=self.R)
+ self.L = self.A * self.KalmanGain
+
+ self.U_max = numpy.matrix([[12.0], [12.0]])
+ self.U_min = numpy.matrix([[-12.0], [-12.0]])
+
+ self.InitializeState()
+
+
+class IntegralArm(Arm):
+ def __init__(self, name="IntegralArm"):
+ super(IntegralArm, self).__init__(name=name)
+
+ self.A_continuous_unaugmented = self.A_continuous
+ self.B_continuous_unaugmented = self.B_continuous
+
+ self.A_continuous = numpy.matrix(numpy.zeros((6, 6)))
+ self.A_continuous[0:4, 0:4] = self.A_continuous_unaugmented
+ self.A_continuous[0:4, 4:6] = self.B_continuous_unaugmented
+
+ self.B_continuous = numpy.matrix(numpy.zeros((6, 2)))
+ self.B_continuous[0:4, 0:2] = self.B_continuous_unaugmented
+
+ self.C_unaugmented = self.C
+ self.C = numpy.matrix(numpy.zeros((2, 6)))
+ self.C[0:2, 0:4] = self.C_unaugmented
+
+ self.A, self.B = self.ContinuousToDiscrete(self.A_continuous, self.B_continuous, self.dt)
+
+ q_pos_shoulder = 0.08
+ q_vel_shoulder = 4.00
+ q_voltage_shoulder = 6.0
+ q_pos_shooter = 0.08
+ q_vel_shooter = 4.00
+ q_voltage_shooter = 6.0
+ self.Q = numpy.matrix(numpy.zeros((6, 6)))
+ self.Q[0, 0] = q_pos_shoulder ** 2.0
+ self.Q[1, 1] = q_vel_shoulder ** 2.0
+ self.Q[2, 2] = q_pos_shooter ** 2.0
+ self.Q[3, 3] = q_vel_shooter ** 2.0
+ self.Q[4, 4] = q_voltage_shoulder ** 2.0
+ self.Q[5, 5] = q_voltage_shooter ** 2.0
+
+ self.R = numpy.matrix(numpy.zeros((2, 2)))
+ r_pos = 0.05
+ self.R[0, 0] = r_pos ** 2.0
+ self.R[1, 1] = r_pos ** 2.0
+
+ self.KalmanGain, self.Q_steady = controls.kalman(
+ A=self.A, B=self.B, C=self.C, Q=self.Q, R=self.R)
+ self.L = self.A * self.KalmanGain
+
+ self.K_unaugmented = self.K
+ self.K = numpy.matrix(numpy.zeros((2, 6)))
+ self.K[0:2, 0:4] = self.K_unaugmented
+ self.K[0, 4] = 1
+ self.K[1, 5] = 1
+
+ self.Kff = numpy.concatenate((self.Kff, numpy.matrix(numpy.zeros((2, 2)))), axis=1)
+
+ self.InitializeState()
+
+
+class ScenarioPlotter(object):
+ def __init__(self):
+ # Various lists for graphing things.
+ self.t = []
+ self.x_shoulder = []
+ self.v_shoulder = []
+ self.a_shoulder = []
+ self.x_hat_shoulder = []
+ self.u_shoulder = []
+ self.offset_shoulder = []
+ self.x_shooter = []
+ self.v_shooter = []
+ self.a_shooter = []
+ self.x_hat_shooter = []
+ self.u_shooter = []
+ self.offset_shooter = []
+ self.goal_x_shoulder = []
+ self.goal_v_shoulder = []
+ self.goal_x_shooter = []
+ self.goal_v_shooter = []
+
+ def run_test(self, arm, end_goal,
+ iterations=200, controller=None, observer=None):
+ """Runs the plant with an initial condition and goal.
+
+ Args:
+ arm: Arm object to use.
+ end_goal: numpy.Matrix[6, 1], end goal state.
+ iterations: Number of timesteps to run the model for.
+ controller: Arm object to get K from, or None if we should
+ use arm.
+ observer: Arm object to use for the observer, or None if we should
+ use the actual state.
+ """
+
+ if controller is None:
+ controller = arm
+
+ vbat = 12.0
+
+ if self.t:
+ initial_t = self.t[-1] + arm.dt
+ else:
+ initial_t = 0
+
+ goal = numpy.concatenate((arm.X, numpy.matrix(numpy.zeros((2, 1)))), axis=0)
+ current_shoulder_goal = goal[0:2, 0].copy()
+ current_shooter_goal = goal[2:4, 0].copy()
+
+ shoulder_profile = TrapizoidProfile(arm.dt)
+ shoulder_profile.set_maximum_acceleration(50.0)
+ shoulder_profile.set_maximum_velocity(10.0)
+ shoulder_profile.SetGoal(current_shoulder_goal[0, 0])
+ shooter_profile = TrapizoidProfile(arm.dt)
+ shooter_profile.set_maximum_acceleration(50.0)
+ shooter_profile.set_maximum_velocity(10.0)
+ shooter_profile.SetGoal(current_shooter_goal[0, 0])
+
+ U_last = numpy.matrix(numpy.zeros((2, 1)))
+ for i in xrange(iterations):
+ X_hat = arm.X
+
+ if observer is not None:
+ observer.Y = arm.Y
+ observer.CorrectObserver(U_last)
+ self.offset_shoulder.append(observer.X_hat[4, 0])
+ self.offset_shooter.append(observer.X_hat[5, 0])
+
+ next_shoulder_goal = shoulder_profile.Update(end_goal[0, 0], end_goal[1, 0])
+ next_shooter_goal = shooter_profile.Update(end_goal[2, 0], end_goal[3, 0])
+
+ next_goal = numpy.concatenate((next_shoulder_goal, next_shooter_goal, numpy.matrix(numpy.zeros((2, 1)))), axis=0)
+ self.goal_x_shoulder.append(goal[0, 0])
+ self.goal_v_shoulder.append(goal[1, 0])
+ self.goal_x_shooter.append(goal[2, 0])
+ self.goal_v_shooter.append(goal[3, 0])
+
+ ff_U = controller.Kff * (next_goal - observer.A * goal)
+
+ if observer is not None:
+ X_hat = observer.X_hat
+ self.x_hat_shoulder.append(observer.X_hat[0, 0])
+ self.x_hat_shooter.append(observer.X_hat[2, 0])
+
+ U_uncapped = controller.K * (goal - X_hat) + ff_U
+ U = U_uncapped.copy()
+
+ U[0, 0] = numpy.clip(U[0, 0], -vbat, vbat)
+ U[1, 0] = numpy.clip(U[1, 0], -vbat, vbat)
+ self.x_shoulder.append(arm.X[0, 0])
+ self.x_shooter.append(arm.X[2, 0])
+
+ if self.v_shoulder:
+ last_v_shoulder = self.v_shoulder[-1]
+ else:
+ last_v_shoulder = 0
+ self.v_shoulder.append(arm.X[1, 0])
+ self.a_shoulder.append(
+ (self.v_shoulder[-1] - last_v_shoulder) / arm.dt)
+
+ if self.v_shooter:
+ last_v_shooter = self.v_shooter[-1]
+ else:
+ last_v_shooter = 0
+ self.v_shooter.append(arm.X[3, 0])
+ self.a_shooter.append(
+ (self.v_shooter[-1] - last_v_shooter) / arm.dt)
+
+ if i % 40 == 0:
+ # Test that if we move the shoulder, the shooter stays perfect.
+ #observer.X_hat[0, 0] += 0.20
+ #arm.X[0, 0] += 0.20
+ pass
+ U_error = numpy.matrix([[0.0], [0.0]])
+ # Kick it and see what happens.
+ #if (initial_t + i * arm.dt) % 0.4 > 0.2:
+ #U_error = numpy.matrix([[4.0], [0.0]])
+ #else:
+ #U_error = numpy.matrix([[-4.0], [0.0]])
+
+ arm.Update(U + U_error)
+
+ if observer is not None:
+ observer.PredictObserver(U)
+
+ self.t.append(initial_t + i * arm.dt)
+ self.u_shoulder.append(U[0, 0])
+ self.u_shooter.append(U[1, 0])
+
+ glog.debug('Time: %f', self.t[-1])
+
+ ff_U -= U_uncapped - U
+ goal = controller.A * goal + controller.B * ff_U
+
+ if U[0, 0] != U_uncapped[0, 0]:
+ glog.debug('Moving shoulder %s', repr(initial_t + i * arm.dt))
+ glog.debug('U error %s', repr(U_uncapped - U))
+ glog.debug('goal change is %s',
+ repr(next_shoulder_goal -
+ numpy.matrix([[goal[0, 0]], [goal[1, 0]]])))
+ shoulder_profile.MoveCurrentState(
+ numpy.matrix([[goal[0, 0]], [goal[1, 0]]]))
+ if U[1, 0] != U_uncapped[1, 0]:
+ glog.debug('Moving shooter %s', repr(initial_t + i * arm.dt))
+ glog.debug('U error %s', repr(U_uncapped - U))
+ shooter_profile.MoveCurrentState(
+ numpy.matrix([[goal[2, 0]], [goal[3, 0]]]))
+ U_last = U
+ glog.debug('End goal is %s', repr(end_goal))
+ glog.debug('last goal is %s', repr(goal))
+ glog.debug('End state is %s', repr(arm.X))
+
+
+ def Plot(self):
+ pylab.subplot(3, 1, 1)
+ pylab.plot(self.t, self.x_shoulder, label='x shoulder')
+ pylab.plot(self.t, self.goal_x_shoulder, label='goal x shoulder')
+ pylab.plot(self.t, self.x_hat_shoulder, label='x_hat shoulder')
+
+ pylab.plot(self.t, self.x_shooter, label='x shooter')
+ pylab.plot(self.t, self.x_hat_shooter, label='x_hat shooter')
+ pylab.plot(self.t, self.goal_x_shooter, label='goal x shooter')
+ pylab.plot(self.t, map(operator.add, self.x_shooter, self.x_shoulder),
+ label='x shooter ground')
+ pylab.plot(self.t, map(operator.add, self.x_hat_shooter, self.x_hat_shoulder),
+ label='x_hat shooter ground')
+ pylab.legend()
+
+ pylab.subplot(3, 1, 2)
+ pylab.plot(self.t, self.u_shoulder, label='u shoulder')
+ pylab.plot(self.t, self.offset_shoulder, label='voltage_offset shoulder')
+ pylab.plot(self.t, self.u_shooter, label='u shooter')
+ pylab.plot(self.t, self.offset_shooter, label='voltage_offset shooter')
+ pylab.legend()
+
+ pylab.subplot(3, 1, 3)
+ pylab.plot(self.t, self.a_shoulder, label='a_shoulder')
+ pylab.plot(self.t, self.a_shooter, label='a_shooter')
+ pylab.legend()
+
+ pylab.show()
+
+
+def main(argv):
+ argv = FLAGS(argv)
+ glog.init()
+
+ scenario_plotter = ScenarioPlotter()
+
+ arm = Arm()
+ arm_controller = IntegralArm()
+ arm_observer = IntegralArm()
+
+ # Test moving the shoulder with constant separation.
+ initial_X = numpy.matrix([[0.0], [0.0], [0.0], [0.0], [0.0], [0.0]])
+ R = numpy.matrix([[numpy.pi / 2.0],
+ [0.0],
+ [0.0], #[numpy.pi / 2.0],
+ [0.0],
+ [0.0],
+ [0.0]])
+ arm.X = initial_X[0:4, 0]
+ arm_observer.X = initial_X
+
+ scenario_plotter.run_test(arm=arm,
+ end_goal=R,
+ iterations=300,
+ controller=arm_controller,
+ observer=arm_observer)
+
+ if len(argv) != 5:
+ glog.fatal('Expected .h file name and .cc file name for the wrist and integral wrist.')
+ else:
+ namespaces = ['y2016', 'control_loops', 'superstructure']
+ loop_writer = control_loop.ControlLoopWriter('Arm', [arm],
+ namespaces=namespaces)
+ loop_writer.Write(argv[1], argv[2])
+
+ integral_loop_writer = control_loop.ControlLoopWriter(
+ 'IntegralArm', [arm_controller], namespaces=namespaces)
+ integral_loop_writer.Write(argv[3], argv[4])
+
+ if FLAGS.plot:
+ scenario_plotter.Plot()
+
+if __name__ == '__main__':
+ sys.exit(main(sys.argv))