def runDircol(self,x0,xf,tf0): N = 15 # constant #N = np.int(tf0 * 10) # "10Hz" / samples per second context = self.CreateDefaultContext() dircol = DirectCollocation(self, context, num_time_samples=N, minimum_timestep=0.05, maximum_timestep=1.0) u = dircol.input() # set some constraints on inputs dircol.AddEqualTimeIntervalsConstraints() dircol.AddConstraintToAllKnotPoints(u[1] <= self.slewmax) dircol.AddConstraintToAllKnotPoints(u[1] >= -self.slewmax) dircol.AddConstraintToAllKnotPoints(u[0] <= self.umax) dircol.AddConstraintToAllKnotPoints(u[0] >= -self.umax) # constrain the last input to be zero (at least for the u input) #import pdb; pdb.set_trace() dv = dircol.decision_variables() for i in range(3, self.nX*N, 4): alfa_state = dv[i] #u[t_end] dircol.AddBoundingBoxConstraint(-self.alfamax, self.alfamax, alfa_state) #final_u_decision_var = dv[self.nX*N + self.nU*N - 1] #u[t_end] #dircol.AddLinearEqualityConstraint(final_u_decision_var, 0.0) #first_u_decision_var = dv[self.nX*N + 1 ] #u[t_0] #dircol.AddLinearEqualityConstraint(first_u_decision_var, 0.0) # set some constraints on start and final pose eps = 0.0 * np.ones(self.nX) # relaxing factor dircol.AddBoundingBoxConstraint(x0, x0, dircol.initial_state()) dircol.AddBoundingBoxConstraint(xf-eps, \ xf+eps, dircol.final_state()) R = 1.0*np.eye(self.nU) # Cost on input "effort". dircol.AddRunningCost( u.transpose().dot(R.dot(u)) ) # Add a final cost equal to the total duration. dircol.AddFinalCost(dircol.time()) # guess initial trajectory initial_x_trajectory = \ PiecewisePolynomial.FirstOrderHold([0., tf0], np.column_stack((x0, xf))) dircol.SetInitialTrajectory(PiecewisePolynomial(), initial_x_trajectory) # optimize result = Solve(dircol) print('******\nRunning trajectory optimization:') print('w/ solver %s' %(result.get_solver_id().name())) print(result.get_solution_result()) assert(result.is_success()) xtraj = dircol.ReconstructStateTrajectory(result) utraj = dircol.ReconstructInputTrajectory(result) # return nominal trajectory return utraj,xtraj
def compute_control(self, x0_p1, x0_p2, xf_p1, xf_p2, obstacles): """This is basically the single-agent MPC algorithm""" prog = DirectCollocation(self.mpc_params.sys_two_players_c, self.mpc_params.sys_two_players_c.CreateDefaultContext(), self.mpc_params.N+1, minimum_timestep=self.mpc_params.minT, maximum_timestep=self.mpc_params.maxT) x0 = np.concatenate((x0_p1, x0_p2), axis=0) prog.AddBoundingBoxConstraint(x0, x0, prog.initial_state()) x_des = np.concatenate((xf_p1, xf_p2)) Q = np.zeros((8,8)) Q[0:4,0:4] = self.mpc_params.Omega_N_max Q[4:8,4:8] = self.mpc_params.Omega_N_max prog.AddQuadraticErrorCost(Q, x_desired=x_des, vars=prog.final_state()) prog.AddEqualTimeIntervalsConstraints() for obs_pos in obstacles: # both players should avoid the other players for n in range(self.mpc_params.N): x = prog.state() prog.AddConstraintToAllKnotPoints((x[0:2]-obs_pos).dot(x[0:2]-obs_pos) >= (2.0*self.sim_params.player_radius)**2) prog.AddConstraintToAllKnotPoints((x[4:6]-obs_pos).dot(x[4:6]-obs_pos) >= (2.0*self.sim_params.player_radius)**2) # players should avoid each other prog.AddConstraintToAllKnotPoints((x[0:2]-x[4:6]).dot(x[0:2]-x[4:6]) >= (2.0*self.sim_params.player_radius)**2) # input constraints for i in range(4): prog.AddConstraintToAllKnotPoints(prog.input()[i] <= self.sim_params.input_limit) prog.AddConstraintToAllKnotPoints(prog.input()[i] >= -self.sim_params.input_limit) r = self.sim_params.player_radius prog.AddConstraintToAllKnotPoints(prog.state()[0] + r <= self.sim_params.arena_limits_x / 2.0) prog.AddConstraintToAllKnotPoints(prog.state()[0] - r >= -self.sim_params.arena_limits_x / 2.0) prog.AddConstraintToAllKnotPoints(prog.state()[1] + r <= self.sim_params.arena_limits_y / 2.0) prog.AddConstraintToAllKnotPoints(prog.state()[1] - r >= -self.sim_params.arena_limits_y / 2.0) prog.AddConstraintToAllKnotPoints(prog.state()[4] + r <= self.sim_params.arena_limits_x / 2.0) prog.AddConstraintToAllKnotPoints(prog.state()[4] - r >= -self.sim_params.arena_limits_x / 2.0) prog.AddConstraintToAllKnotPoints(prog.state()[5] + r <= self.sim_params.arena_limits_y / 2.0) prog.AddConstraintToAllKnotPoints(prog.state()[5] - r >= -self.sim_params.arena_limits_y / 2.0) prog.AddFinalCost(prog.time()) if not self.prev_u is None and not self.prev_x is None: prog.SetInitialTrajectory(traj_init_u=self.prev_u, traj_init_x=self.prev_x) solver = SnoptSolver() result = solver.Solve(prog) u_traj = prog.ReconstructInputTrajectory(result) x_traj = prog.ReconstructStateTrajectory(result) self.prev_u = u_traj self.prev_x = x_traj u_vals = u_traj.vector_values(u_traj.get_segment_times()) x_vals = x_traj.vector_values(x_traj.get_segment_times()) return True, u_vals[0:2,0], u_vals[2:4,0]
def compute_control(self, x_des, sim_state, team_name, player_id): prog = DirectCollocation(self.mpc_params.sys, self.mpc_params.sys.CreateDefaultContext(), self.mpc_params.N, minimum_timestep=self.mpc_params.minT, maximum_timestep=self.mpc_params.maxT) pos0 = sim_state.get_player_pos(team_name, player_id) vel0 = sim_state.get_player_vel(team_name, player_id) x0 = np.concatenate((pos0, vel0), axis=0) prog.AddBoundingBoxConstraint(x0, x0, prog.initial_state()) prog.AddQuadraticErrorCost(Q=self.mpc_params.Omega_N_max, x_desired=x_des, vars=prog.final_state()) obstacle_positions = self.get_obstacle_positions( sim_state, team_name, player_id) for obs_pos in obstacle_positions: for n in range(self.mpc_params.N): x = prog.state() prog.AddConstraintToAllKnotPoints( (x[0:2] - obs_pos).dot(x[0:2] - obs_pos) >= ( 2.0 * self.sim_params.player_radius)**2) prog.AddEqualTimeIntervalsConstraints() self.add_input_limits(prog) self.add_arena_limits(prog) prog.AddFinalCost(prog.time()) if not self.prev_u is None and not self.prev_x is None: prog.SetInitialTrajectory(traj_init_u=self.prev_u, traj_init_x=self.prev_x) solver = SnoptSolver() result = solver.Solve(prog) u_traj = prog.ReconstructInputTrajectory(result) x_traj = prog.ReconstructStateTrajectory(result) u_vals = u_traj.vector_values(u_traj.get_segment_times()) self.prev_u = u_traj self.prev_x = x_traj return u_vals[:, 0]
def runDircol(self, x0, xf, tf0): N = 21 #np.int(tf0 * 10) # "10Hz" samples per second context = self.CreateDefaultContext() dircol = DirectCollocation(self, context, num_time_samples=N, minimum_timestep=0.05, maximum_timestep=1.0) u = dircol.input() dircol.AddEqualTimeIntervalsConstraints() dircol.AddConstraintToAllKnotPoints(u[0] <= 0.5 * self.omegamax) dircol.AddConstraintToAllKnotPoints(u[0] >= -0.5 * self.omegamax) dircol.AddConstraintToAllKnotPoints(u[1] <= 0.5 * self.umax) dircol.AddConstraintToAllKnotPoints(u[1] >= -0.5 * self.umax) eps = 0.0 dircol.AddBoundingBoxConstraint(x0, x0, dircol.initial_state()) dircol.AddBoundingBoxConstraint(xf - np.array([eps, eps, eps]), xf + np.array([eps, eps, eps]), dircol.final_state()) R = 1.0 * np.eye(2) # Cost on input "effort". dircol.AddRunningCost(u.transpose().dot(R.dot(u))) #dircol.AddRunningCost(R*u[0]**2) # Add a final cost equal to the total duration. dircol.AddFinalCost(dircol.time()) initial_x_trajectory = \ PiecewisePolynomial.FirstOrderHold([0., tf0], np.column_stack((x0, xf))) dircol.SetInitialTrajectory(PiecewisePolynomial(), initial_x_trajectory) result = Solve(dircol) print(result.get_solver_id().name()) print(result.get_solution_result()) assert (result.is_success()) #import pdb; pdb.set_trace() xtraj = dircol.ReconstructStateTrajectory(result) utraj = dircol.ReconstructInputTrajectory(result) return utraj, xtraj
def make_real_dircol_mp(): global tree global plant global context global dircol # expmt = "acrobot" # State: (theta1, theta2, theta1_dot, theta2_dot) Input: Elbow torque tree = RigidBodyTree("/opt/underactuated/src/acrobot/acrobot.urdf", FloatingBaseType.kFixed) plant = AcrobotPlant() context = plant.CreateDefaultContext() dircol = DirectCollocation(plant, context, num_time_samples=21, minimum_timestep=0.05, maximum_timestep=0.2) dircol.AddEqualTimeIntervalsConstraints() # Add input limits. torque_limit = 8.0 # N*m. u = dircol.input() dircol.AddConstraintToAllKnotPoints(-torque_limit <= u[0]) dircol.AddConstraintToAllKnotPoints(u[0] <= torque_limit) initial_state = (0., 0., 0., 0.) dircol.AddBoundingBoxConstraint(initial_state, initial_state, dircol.initial_state()) final_state = (math.pi, 0., 0., 0.) dircol.AddBoundingBoxConstraint(final_state, final_state, dircol.final_state()) R = 10 # Cost on input "effort". u = dircol.input() dircol.AddRunningCost(R*u[0]**2) # Add a final cost equal to the total duration. dircol.AddFinalCost(dircol.time()) initial_x_trajectory = \ PiecewisePolynomial.FirstOrderHold([0., 4.], np.column_stack((initial_state, final_state))) dircol.SetInitialTrajectory(PiecewisePolynomial(), initial_x_trajectory) print(id(dircol)) return dircol
# Initial state on the surface of section (and velocity > .1). dircol.AddBoundingBoxConstraint([0., 0.1], [0., 10.], dircol.initial_state()) # Periodicity constraint. # TODO(russt): Replace this with the vectorized version pending drake #8315. dircol.AddLinearConstraint( dircol.final_state()[0] == dircol.initial_state()[0]) dircol.AddLinearConstraint( dircol.final_state()[1] == dircol.initial_state()[1]) samples = np.linspace(0, 2*math.pi, 10) x_guess = np.vstack(([2*math.sin(t) for t in samples], [2*math.cos(t) for t in samples])) initial_x_trajectory = PiecewisePolynomial.FirstOrderHold(samples, x_guess) dircol.SetInitialTrajectory(PiecewisePolynomial(), initial_x_trajectory) fig = plt.figure() h, = plt.plot([], [], '.-') plt.xlim((-2.5, 2.5)) plt.ylim((-3., 3.)) def draw_trajectory(t, x): h.set_xdata(x[0, :]) h.set_ydata(x[1, :]) fig.canvas.draw() if plt.get_backend() == u'MacOSX': plt.pause(1e-10)
def drake_trajectory_generation(file_name): global x_cmd_drake global u_cmd_drake print(file_name) Parser(plant).AddModelFromFile(file_name) plant.Finalize() context = plant.CreateDefaultContext() global dircol dircol= DirectCollocation( plant, context, num_time_samples=11, minimum_timestep=0.1, maximum_timestep=0.4, input_port_index=plant.get_actuation_input_port().get_index()) dircol.AddEqualTimeIntervalsConstraints() initial_state = (0., 0., 0., 0.) dircol.AddBoundingBoxConstraint(initial_state, initial_state, dircol.initial_state()) final_state = (0., math.pi, 0., 0.) dircol.AddBoundingBoxConstraint(final_state, final_state, dircol.final_state()) R = 10 # Cost on input "effort".weight u = dircol.input() dircol.AddRunningCost(R * u[0]**2) # Add a final cost equal to the total duration. dircol.AddFinalCost(dircol.time()) initial_x_trajectory = PiecewisePolynomial.FirstOrderHold( [0., 4.], np.column_stack((initial_state, final_state))) # yapf: disable dircol.SetInitialTrajectory(PiecewisePolynomial(), initial_x_trajectory) dircol.AddConstraintToAllKnotPoints(dircol.input()[1] <= 0) dircol.AddConstraintToAllKnotPoints(dircol.input()[1] >= 0) global result global u_values result = Solve(dircol) assert result.is_success() #plotphase_portrait() fig1, ax1 = plt.subplots() u_trajectory = dircol.ReconstructInputTrajectory(result) u_knots = np.hstack([ u_trajectory.value(t) for t in np.linspace(u_trajectory.start_time(), u_trajectory.end_time(), 400) ])#here the u_knots now is 2x400 #u_trajectory = dircol.ReconstructInputTrajectory(result) times = np.linspace(u_trajectory.start_time(), u_trajectory.end_time(), 400) #u_lookup = np.vectorize(u_trajectory.value) #now we have ndarray of u_values with 400 points for 4 seconds w/ 100hz pub frequency #u_values = u_lookup(times) #ax1.plot(times, u_values) ax1.plot(times, u_knots[0]) ax1.plot(times, u_knots[1]) ax1.set_xlabel("time (seconds)") ax1.set_ylabel("force (Newtons)") ax1.set_title(' Direct collocation for Cartpole ') print('here2') plt.show() print('here3') #x_knots = np.hstack([ # x_trajectory.value(t) for t in np.linspace(x_trajectory.start_time(), # x_trajectory.end_time(), 100) #]) x_trajectory = dircol.ReconstructStateTrajectory(result) x_knots = np.hstack([ x_trajectory.value(t) for t in np.linspace(x_trajectory.start_time(), x_trajectory.end_time(), 400) ]) print(x_trajectory.start_time()) print(x_trajectory.end_time()) fig, ax = plt.subplots(4,1,figsize=(8,8)) plt.subplots_adjust(wspace =0, hspace =0.4) #plt.tight_layout(3)#adjust total space ax[0].set_title('state of direct collocation for Cartpole') ax[0].plot(x_knots[0, :], x_knots[2, :], linewidth=2, color='b', linestyle='-') ax[0].set_xlabel("state_dot(theta1_dot and t|heta2_dot)") ax[0].set_ylabel("state(theta1 and theta2)"); ax[0].plot(x_knots[1, :], x_knots[3, :],color='r',linewidth=2,linestyle='--') ax[0].legend(('theta1&theta1dot','theta2&theta2dot')); ax[1].set_title('input u(t) of direct collocation for Cartpole') # ax[1].plot(times,u_values, 'g') ax[1].plot(times, u_knots[0]) ax[1].plot(times, u_knots[1]) ax[1].legend(('input u(t)')) ax[1].set_xlabel("time") ax[1].set_ylabel("u(t)") ax[1].legend(('x joint ','thetajoint')) ax[1].set_title('input x(t) of direct collocation for Cartpole') ax[2].plot(times, x_knots[0, :]) ax[2].set_xlabel("time") ax[2].set_ylabel("x(t)") ax[2].set_title('input theta(t) of direct collocation for Cartpole') ax[3].set_title('input theta(t) of direct collocation for Cartpole') ax[3].plot(times, x_knots[1, :]) ax[3].set_xlabel("time") ax[3].set_ylabel("theta(t)") print('here4') plt.show() print('here5') x_cmd_drake=x_knots #return x_knots[0, :]#u_values # u_cmd_drake=u_values u_cmd_drake=u_knots
def make_dircol_pendulum(ic=(-1., 0.), num_samples=32, min_timestep=0.002, max_timestep=0.25, warm_start="linear", seed=1776, should_vis=False, target_traj=None, **kwargs): # if 'warm_start' in kwargs: # print(kwargs['warm_start']) # else: # print("warm_start", warm_start) global dircol global plant global context plant = PendulumPlant() context = plant.CreateDefaultContext() dircol = DirectCollocation(plant, context, num_time_samples=num_samples, minimum_timestep=min_timestep, maximum_timestep=max_timestep) dircol.AddEqualTimeIntervalsConstraints() # torque_limit = input_limit # N*m. torque_limit = 5. u = dircol.input() dircol.AddConstraintToAllKnotPoints(-torque_limit <= u[0]) dircol.AddConstraintToAllKnotPoints(u[0] <= torque_limit) initial_state = ic dircol.AddBoundingBoxConstraint(initial_state, initial_state, dircol.initial_state()) final_state = (math.pi, 0.) dircol.AddBoundingBoxConstraint(final_state, final_state, dircol.final_state()) # R = 100 # Cost on input "effort". u = dircol.input() x = dircol.state() # print(x) dircol.AddRunningCost(2 * ((x[0] - math.pi) * (x[0] - math.pi) + x[1] * x[1]) + 25 * u.dot(u)) # Add a final cost equal to the total duration. # dircol.AddFinalCost(dircol.time()) if warm_start == "linear": initial_u_trajectory = PiecewisePolynomial() initial_x_trajectory = \ PiecewisePolynomial.FirstOrderHold([0., 4.], np.column_stack((initial_state, final_state))) dircol.SetInitialTrajectory(initial_u_trajectory, initial_x_trajectory) elif warm_start == "random": assert isinstance(seed, int) np.random.seed(seed) breaks = np.linspace(0, 4, num_samples).reshape( (-1, 1)) # using num_time_samples u_knots = np.random.rand( 1, num_samples) - 0.5 # num_inputs vs num_samples? x_knots = np.random.rand( 2, num_samples) - 0.5 # num_states vs num_samples? initial_u_trajectory = PiecewisePolynomial.Cubic( breaks, u_knots, False) initial_x_trajectory = PiecewisePolynomial.Cubic( breaks, x_knots, False) dircol.SetInitialTrajectory(initial_u_trajectory, initial_x_trajectory) elif warm_start == "target": assert target_traj != [], "Need a valid target for warm starting" (breaks, x_knots, u_knots) = target_traj #(breaks, u_knots, x_knots) = target_traj initial_u_trajectory = PiecewisePolynomial.Cubic( breaks.T, u_knots.T, False) initial_x_trajectory = PiecewisePolynomial.Cubic( breaks.T, x_knots.T, False) dircol.SetInitialTrajectory(initial_u_trajectory, initial_x_trajectory) def cb(decision_vars): global vis_cb_counter vis_cb_counter += 1 if vis_cb_counter % 10 != 0: return # Get the total cost all_costs = dircol.EvalBindings(dircol.GetAllCosts(), decision_vars) # Get the total cost of the constraints. # Additionally, the number and extent of any constraint violations. violated_constraint_count = 0 violated_constraint_cost = 0 constraint_cost = 0 for constraint in dircol.GetAllConstraints(): val = dircol.EvalBinding(constraint, decision_vars) # Consider switching to DoCheckSatisfied if you can find the binding... nudge = 1e-1 # This much constraint violation is not considered bad... lb = constraint.evaluator().lower_bound() ub = constraint.evaluator().upper_bound() good_lb = np.all(np.less_equal(lb, val + nudge)) good_ub = np.all(np.greater_equal(ub, val - nudge)) if not good_lb or not good_ub: # print("{} <= {} <= {}".format(lb, val, ub)) violated_constraint_count += 1 # violated_constraint_cost += np.sum(np.abs(val)) if not good_lb: violated_constraint_cost += np.sum(np.abs(lb - val)) if not good_ub: violated_constraint_cost += np.sum(np.abs(val - ub)) constraint_cost += np.sum(np.abs(val)) print("total cost: {: .2f} | \tconstraint {: .2f} \tbad {}, {: .2f}". format(sum(all_costs), constraint_cost, violated_constraint_count, violated_constraint_cost)) #dircol.AddVisualizationCallback(cb, dircol.decision_variables()) def MyVisualization(sample_times, values): global vis_cb_counter vis_cb_counter += 1 #print("counter: ", vis_cb_counter) if vis_cb_counter % 10 != 0: return x, x_dot = values[0], values[1] plt.plot(x, x_dot, '-o', label=vis_cb_counter) plt.show() if should_vis: plt.figure() plt.title('Tip trajectories') plt.xlabel('x') plt.ylabel('x_dot') dircol.AddStateTrajectoryCallback(MyVisualization) from pydrake.all import (SolverType) #dircol.SetSolverOption(SolverType.kSnopt, 'Major feasibility tolerance', 1.0e-6) # default="1.0e-6" #dircol.SetSolverOption(SolverType.kSnopt, 'Major optimality tolerance', 1.0e-6) # default="1.0e-6" #dircol.SetSolverOption(SolverType.kSnopt, 'Minor feasibility tolerance', 1.0e-6) # default="1.0e-6" #dircol.SetSolverOption(SolverType.kSnopt, 'Minor optimality tolerance', 1.0e-6) # default="1.0e-6" # dircol.SetSolverOption(SolverType.kSnopt, 'Major feasibility tolerance', 1.0e-6) # default="1.0e-6" # dircol.SetSolverOption(SolverType.kSnopt, 'Major optimality tolerance', 5.0e-2) # default="1.0e-6" was 5.0e-1 # dircol.SetSolverOption(SolverType.kSnopt, 'Minor feasibility tolerance', 1.0e-6) # default="1.0e-6" # dircol.SetSolverOption(SolverType.kSnopt, 'Minor optimality tolerance', 5.0e-2) # default="1.0e-6" was 5.0e-1 dircol.SetSolverOption( SolverType.kSnopt, 'Time limit (secs)', 60.0) # default="9999999.0" # Very aggressive cutoff... dircol.SetSolverOption( SolverType.kSnopt, 'Major step limit', 0.1) # default="2.0e+0" # HUGE!!! default takes WAY too huge steps # dircol.SetSolverOption(SolverType.kSnopt, 'Reduced Hessian dimension', 10000) # Default="min{2000, n1 + 1}" # dircol.SetSolverOption(SolverType.kSnopt, 'Hessian updates', 30) # Default="10" dircol.SetSolverOption(SolverType.kSnopt, 'Major iterations limit', 9300000) # Default="9300" dircol.SetSolverOption(SolverType.kSnopt, 'Minor iterations limit', 50000) # Default="500" dircol.SetSolverOption(SolverType.kSnopt, 'Iterations limit', 50 * 10000) # Default="10000" # Factoriztion? # dircol.SetSolverOption(SolverType.kSnopt, 'QPSolver Cholesky', True) # Default="*Cholesky/CG/QN" #dircol.SetSolverOption(SolverType.kSnopt, 'Major iterations limit', 1) # Default="9300" #dircol.SetSolverOption(SolverType.kSnopt, 'Minor iterations limit', 1) # Default="500" return dircol
def make_dircol_cartpole(ic=(-1., 0., 0., 0.), num_samples=21, min_timestep=0.0001, max_timestep=1., warm_start="linear", seed=1776, should_vis=False, torque_limit=250., target_traj=None, **kwargs): global dircol global plant global context tree = RigidBodyTree("/opt/underactuated/src/cartpole/cartpole.urdf", FloatingBaseType.kFixed) plant = RigidBodyPlant(tree) context = plant.CreateDefaultContext() dircol = DirectCollocation( plant, context, num_time_samples=num_samples, # minimum_timestep=0.01, maximum_timestep=0.01) minimum_timestep=min_timestep, maximum_timestep=max_timestep) # dircol.AddEqualTimeIntervalsConstraints() # torque_limit = input_limit # N*m. # torque_limit = 64. u = dircol.input() dircol.AddConstraintToAllKnotPoints(-torque_limit <= u[0]) dircol.AddConstraintToAllKnotPoints(u[0] <= torque_limit) initial_state = ic dircol.AddBoundingBoxConstraint(initial_state, initial_state, dircol.initial_state()) final_state = np.array([0., math.pi, 0., 0.]).astype(np.double) dircol.AddBoundingBoxConstraint(final_state, final_state, dircol.final_state()) # R = 100 # Cost on input "effort". u = dircol.input() x = dircol.state() # t = dircol.time() # let's add 100*t (seconds) to get in that min-time component! denom1 = float(10**2 + math.pi**2 + 10**2 + math.pi**2) denom2 = float(180**2) #denom1 = 10**2+math.pi**2+10**2+math.pi**2 #denom2 = 180**2 # dircol.AddRunningCost(u.dot(u)/denom2) # dircol.AddRunningCost(2*(x-final_state).dot(x-final_state)/denom1) dircol.AddRunningCost(1 + 2. * (x - final_state).dot(x - final_state) / denom1 + u.dot(u) / denom2) # Add a final cost equal to the total duration. #dircol.AddFinalCost(dircol.time()) # Enabled to sim min time cost? if warm_start == "linear": initial_u_trajectory = PiecewisePolynomial() initial_x_trajectory = \ PiecewisePolynomial.FirstOrderHold([0., 4.], np.column_stack((initial_state, final_state))) dircol.SetInitialTrajectory(initial_u_trajectory, initial_x_trajectory) elif warm_start == "random": assert isinstance(seed, int) np.random.seed(seed) breaks = np.linspace(0, 4, num_samples).reshape( (-1, 1)) # using num_time_samples u_knots = np.random.rand( 1, num_samples) - 0.5 # num_inputs vs num_samples? x_knots = np.random.rand( 2, num_samples) - 0.5 # num_states vs num_samples? initial_u_trajectory = PiecewisePolynomial.Cubic( breaks, u_knots, False) initial_x_trajectory = PiecewisePolynomial.Cubic( breaks, x_knots, False) dircol.SetInitialTrajectory(initial_u_trajectory, initial_x_trajectory) elif warm_start == "target": assert target_traj != [], "Need a valid target for warm starting" (breaks, x_knots, u_knots) = target_traj #(breaks, u_knots, x_knots) = target_traj initial_u_trajectory = PiecewisePolynomial.Cubic( breaks.T, u_knots.T, False) initial_x_trajectory = PiecewisePolynomial.Cubic( breaks.T, x_knots.T, False) dircol.SetInitialTrajectory(initial_u_trajectory, initial_x_trajectory) def cb(decision_vars): global vis_cb_counter vis_cb_counter += 1 if vis_cb_counter % 10 != 0: return # Get the total cost all_costs = dircol.EvalBindings(dircol.GetAllCosts(), decision_vars) # Get the total cost of the constraints. # Additionally, the number and extent of any constraint violations. violated_constraint_count = 0 violated_constraint_cost = 0 constraint_cost = 0 for constraint in dircol.GetAllConstraints(): val = dircol.EvalBinding(constraint, decision_vars) # Consider switching to DoCheckSatisfied if you can find the binding... nudge = 1e-1 # This much constraint violation is not considered bad... lb = constraint.evaluator().lower_bound() ub = constraint.evaluator().upper_bound() good_lb = np.all(np.less_equal(lb, val + nudge)) good_ub = np.all(np.greater_equal(ub, val - nudge)) if not good_lb or not good_ub: # print("{} <= {} <= {}".format(lb, val, ub)) violated_constraint_count += 1 # violated_constraint_cost += np.sum(np.abs(val)) if not good_lb: violated_constraint_cost += np.sum(np.abs(lb - val)) if not good_ub: violated_constraint_cost += np.sum(np.abs(val - ub)) constraint_cost += np.sum(np.abs(val)) print("total cost: {: .2f} | \tconstraint {: .2f} \tbad {}, {: .2f}". format(sum(all_costs), constraint_cost, violated_constraint_count, violated_constraint_cost)) #dircol.AddVisualizationCallback(cb, dircol.decision_variables()) def MyVisualization(sample_times, values): def state_to_tip_coord(state_vec): # State: (x, theta, x_dot, theta_dot) x, theta, _, _ = state_vec pole_length = 0.5 # manually looked this up #return (x-pole_length*np.sin(theta), pole_length-np.cos(theta)) return (x - pole_length * np.sin(theta), pole_length * (-np.cos(theta))) global vis_cb_counter vis_cb_counter += 1 if vis_cb_counter % 30 != 0: return coords = [state_to_tip_coord(state) for state in values.T] x, y = zip(*coords) plt.plot(x, y, '-o', label=vis_cb_counter) #plt.show() # good? #if should_vis: if False: plt.figure() plt.title('Tip trajectories') plt.xlabel('x') plt.ylabel('x_dot') dircol.AddStateTrajectoryCallback(MyVisualization) from pydrake.all import (SolverType) # dircol.SetSolverOption(SolverType.kSnopt, 'Major feasibility tolerance', 1.0e-6) # default="1.0e-6" # dircol.SetSolverOption(SolverType.kSnopt, 'Major optimality tolerance', 5.0e-2) # default="1.0e-6" was 5.0e-1 # dircol.SetSolverOption(SolverType.kSnopt, 'Minor feasibility tolerance', 1.0e-6) # default="1.0e-6" # dircol.SetSolverOption(SolverType.kSnopt, 'Minor optimality tolerance', 5.0e-2) # default="1.0e-6" was 5.0e-1 # dircol.SetSolverOption(SolverType.kSnopt, 'Time limit (secs)', 12.0) # default="9999999.0" # Very aggressive cutoff... dircol.SetSolverOption( SolverType.kSnopt, 'Major step limit', 0.1) # default="2.0e+0" # HUGE!!! default takes WAY too huge steps dircol.SetSolverOption(SolverType.kSnopt, 'Time limit (secs)', 15.0) # default="9999999.0" # was 15 # dircol.SetSolverOption(SolverType.kSnopt, 'Reduced Hessian dimension', 10000) # Default="min{2000, n1 + 1}" # dircol.SetSolverOption(SolverType.kSnopt, 'Hessian updates', 30) # Default="10" dircol.SetSolverOption(SolverType.kSnopt, 'Major iterations limit', 9300000) # Default="9300" dircol.SetSolverOption(SolverType.kSnopt, 'Minor iterations limit', 50000) # Default="500" dircol.SetSolverOption(SolverType.kSnopt, 'Iterations limit', 50 * 10000) # Default="10000" # Factoriztion? # dircol.SetSolverOption(SolverType.kSnopt, 'QPSolver Cholesky', True) # Default="*Cholesky/CG/QN" return dircol
def get_initial_guess(self, x_p1, p_goal, p_puck, obstacles): """This is basically the single-agent MPC algorithm""" hit_dir = p_goal - p_puck hit_dir = 6.0 * hit_dir / np.linalg.norm(hit_dir) x_des = np.array([p_puck[0], p_puck[1], hit_dir[0], hit_dir[1]]) #x_des = np.array([1.0, 1.0, 0, 0]) print("x_des: {}, {}".format(x_des[0], x_des[1])) print("x_des shape", x_des.shape) print("zeros.shape", np.zeros(4).shape) print("p_player", x_p1[0:2]) print("p_puck {}, {}".format(p_puck[0], p_puck[1])) print("p_goal", p_goal) prog = DirectCollocation(self.mpc_params.sys_c, self.mpc_params.sys_c.CreateDefaultContext(), self.mpc_params.N + 1, minimum_timestep=self.mpc_params.minT, maximum_timestep=self.mpc_params.maxT) prog.AddBoundingBoxConstraint(x_p1, x_p1, prog.initial_state()) prog.AddQuadraticErrorCost(Q=self.mpc_params.Omega_N_max, x_desired=x_des, vars=prog.final_state()) prog.AddEqualTimeIntervalsConstraints() # generate trajectory non in collision with puck #for n in range(self.mpc_params.N): # x = prog.state() # eps = 0.1 # obs_pos = p_puck[0:2] # prog.AddConstraintToAllKnotPoints((x[0:2]-obs_pos).dot(x[0:2]-obs_pos) >= (self.sim_params.player_radius + self.sim_params.puck_radius - eps)**2) for obs_pos in obstacles: for n in range(self.mpc_params.N): x = prog.state() prog.AddConstraintToAllKnotPoints( (x[0:2] - obs_pos).dot(x[0:2] - obs_pos) >= ( 2.0 * self.sim_params.player_radius)**2) prog.AddConstraintToAllKnotPoints( prog.input()[0] <= self.sim_params.input_limit) prog.AddConstraintToAllKnotPoints( prog.input()[0] >= -self.sim_params.input_limit) prog.AddConstraintToAllKnotPoints( prog.input()[1] <= self.sim_params.input_limit) prog.AddConstraintToAllKnotPoints( prog.input()[1] >= -self.sim_params.input_limit) r = self.sim_params.player_radius prog.AddConstraintToAllKnotPoints( prog.state()[0] + r <= self.sim_params.arena_limits_x / 2.0) prog.AddConstraintToAllKnotPoints( prog.state()[0] - r >= -self.sim_params.arena_limits_x / 2.0) prog.AddConstraintToAllKnotPoints( prog.state()[1] + r <= self.sim_params.arena_limits_y / 2.0) prog.AddConstraintToAllKnotPoints( prog.state()[1] - r >= -self.sim_params.arena_limits_y / 2.0) prog.AddFinalCost(prog.time()) if not self.prev_u is None and not self.prev_x is None: prog.SetInitialTrajectory(traj_init_u=self.prev_u, traj_init_x=self.prev_x) solver = SnoptSolver() result = solver.Solve(prog) u_traj = prog.ReconstructInputTrajectory(result) x_traj = prog.ReconstructStateTrajectory(result) self.prev_u = u_traj self.prev_x = x_traj u_vals = u_traj.vector_values(u_traj.get_segment_times()) x_vals = x_traj.vector_values(x_traj.get_segment_times()) print(u_vals) print(u_vals[:, 0]) return u_vals[:, 0]
# Add initial and final state constraints x0 = [0, 0, 0, 0] xf = [pi, 0, 0, 0] prog.AddBoundingBoxConstraint(x0, x0, prog.initial_state()) prog.AddBoundingBoxConstraint(xf, xf, prog.final_state()) # Add running cost u = prog.input() R = 10 prog.AddRunningCost(R * u[0]**2) # Add final cost prog.AddFinalCost(prog.time()) # Create an initial guess at the trajectory init_x = PiecewisePolynomial.FirstOrderHold([0, 10.0], np.column_stack( (x0, xf))) init_u = PiecewisePolynomial.FirstOrderHold([0, 10.0], np.zeros(shape=(1, 2))) prog.SetInitialTrajectory(traj_init_u=init_u, traj_init_x=init_x) # Set SNOPT options prog.SetSolverOption(SnoptSolver().solver_id(), "Major Iterations Limit", 1000) prog.SetSolverOption(SnoptSolver().solver_id(), "Major Feasibility Tolerance", 1e-4) prog.SetSolverOption(SnoptSolver().solver_id(), "Major Optimality Tolerance", 1e-4) prog.SetSolverOption(SnoptSolver().solver_id(), "Scale Option", 2) solver = SnoptSolver() # Solve the problem print("Solving trajectory optimization") start = timeit.default_timer() result = solver.Solve(prog) stop = timeit.default_timer() print("Elapsed Time: ", stop - start) # Get the details of the solution
def do_single_basic_pend_traj_opt(should_init): plant = PendulumPlant() context = plant.CreateDefaultContext() kNumTimeSamples = 15 kMinimumTimeStep = 0.01 kMaximumTimeStep = 0.2 dircol = DirectCollocation(plant, context, kNumTimeSamples, kMinimumTimeStep, kMaximumTimeStep) dircol.AddEqualTimeIntervalsConstraints() kTorqueLimit = 3.0 # N*m. u = dircol.input() dircol.AddConstraintToAllKnotPoints(-kTorqueLimit <= u[0]) dircol.AddConstraintToAllKnotPoints(u[0] <= kTorqueLimit) initial_state = PendulumState() initial_state.set_theta(0.0) initial_state.set_thetadot(0.0) dircol.AddBoundingBoxConstraint(initial_state.get_value(), initial_state.get_value(), dircol.initial_state()) # dircol.AddLinearConstraint(dircol.initial_state() == initial_state.get_value()) final_state = PendulumState() final_state.set_theta(math.pi) final_state.set_thetadot(0.0) dircol.AddBoundingBoxConstraint(final_state.get_value(), final_state.get_value(), dircol.final_state()) # dircol.AddLinearConstraint(dircol.final_state() == final_state.get_value()) R = 10 # Cost on input "effort". dircol.AddRunningCost(R*u[0]**2) if should_init: initial_x_trajectory = PiecewisePolynomial.FirstOrderHold([0., 4.], [initial_state.get_value(), final_state.get_value()]) dircol.SetInitialTrajectory(PiecewisePolynomial(), initial_x_trajectory) def cb(decision_vars): global cb_counter cb_counter += 1 if cb_counter % 10 != 1: return # Get the total cost all_costs = dircol.EvalBindings(dircol.GetAllCosts(), decision_vars) # :all_constraints = dircol.EvalBindings(dircol.GetAllConstraints(), decision_vars) # Get the total cost of the constraints. # Additionally, the number and extent of any constraint violations. violated_constraint_count = 0 violated_constraint_cost = 0 constraint_cost = 0 for constraint in dircol.GetAllConstraints(): val = dircol.EvalBinding(constraint, decision_vars) # TODO: switch to DoCheckSatisfied... nudge = 1e-1 # This much constraint violation is not considered bad... lb = constraint.evaluator().lower_bound() ub = constraint.evaluator().upper_bound() good_lb = np.all( np.less_equal(lb, val+nudge) ) good_ub = np.all( np.greater_equal(ub, val-nudge) ) if not good_lb or not good_ub: # print("val,lb,ub: ", val, lb, ub) violated_constraint_count += 1 violated_constraint_cost += np.sum(val) constraint_cost += np.sum(val) print("total cost: {:.2f} | constraint {:.2f}, bad {}, {:.2f}".format( sum(all_costs), constraint_cost, violated_constraint_count, violated_constraint_cost)) # dircol.AddVisualizationCallback(cb, np.array(dircol.decision_variables())) result = dircol.Solve() assert(result == SolutionResult.kSolutionFound) sol_costs = np.hstack([dircol.EvalBindingAtSolution(cost) for cost in dircol.GetAllCosts()]) sol_constraints = np.hstack([dircol.EvalBindingAtSolution(constraint) for constraint in dircol.GetAllConstraints()])
def direct_collocation_zhao_glider(): print("Running direct collocation") plant = SlotineGlider() context = plant.CreateDefaultContext() N = 21 initial_guess = True max_dt = 0.5 max_tf = N * max_dt dircol = DirectCollocation( plant, context, num_time_samples=N, minimum_timestep=0.05, maximum_timestep=max_dt, ) # Constrain all timesteps, $h[k]$, to be equal, so the trajectory breaks are evenly distributed. dircol.AddEqualTimeIntervalsConstraints() # Add input constraints u = dircol.input() dircol.AddConstraintToAllKnotPoints(0 <= u[0]) dircol.AddConstraintToAllKnotPoints(u[0] <= 3) dircol.AddConstraintToAllKnotPoints(-np.pi / 2 <= u[1]) dircol.AddConstraintToAllKnotPoints(u[1] <= np.pi / 2) # Add state constraints x = dircol.state() min_speed = 5 dircol.AddConstraintToAllKnotPoints(x[0] >= min_speed) min_height = 0.5 dircol.AddConstraintToAllKnotPoints(x[3] >= min_height) # Add initial state travel_angle = (3 / 2) * np.pi h0 = 10 dir_vector = np.array([np.cos(travel_angle), np.sin(travel_angle)]) # Start at initial position x0_pos = np.array([h0, 0, 0]) dircol.AddBoundingBoxConstraint(x0_pos, x0_pos, dircol.initial_state()[3:6]) # Periodicity constraints dircol.AddLinearConstraint( dircol.final_state()[0] == dircol.initial_state()[0]) dircol.AddLinearConstraint( dircol.final_state()[1] == dircol.initial_state()[1]) dircol.AddLinearConstraint( dircol.final_state()[2] == dircol.initial_state()[2]) dircol.AddLinearConstraint( dircol.final_state()[3] == dircol.initial_state()[3]) # Always end in right direction # NOTE this assumes that we always are starting in origin if travel_angle % np.pi == 0: # Travel along x-axis dircol.AddConstraint( dircol.final_state()[5] == dircol.initial_state()[5]) elif travel_angle % ((1 / 2) * np.pi) == 0: # Travel along y-axis dircol.AddConstraint( dircol.final_state()[4] == dircol.initial_state()[4]) else: dircol.AddConstraint( dircol.final_state()[5] == dircol.final_state()[4] * np.tan(travel_angle)) # Maximize distance travelled in desired direction p0 = dircol.initial_state() p1 = dircol.final_state() Q = 1 dist_travelled = np.array([p1[4], p1[5]]) # NOTE assume starting in origin dircol.AddFinalCost(-(dir_vector.T.dot(dist_travelled)) * Q) if True: # Cost on input effort R = 0.1 dircol.AddRunningCost(R * (u[0])**2 + R * u[1]**2) # Initial guess is a straight line from x0 in direction if initial_guess: avg_vel_guess = 10 # Guess for initial velocity x0_guess = np.array([avg_vel_guess, travel_angle, 0, h0, 0, 0]) guessed_total_dist_travelled = 200 xf_guess = np.array([ avg_vel_guess, travel_angle, 0, h0, dir_vector[0] * guessed_total_dist_travelled, dir_vector[1] * guessed_total_dist_travelled, ]) initial_x_trajectory = PiecewisePolynomial.FirstOrderHold( [0.0, 4.0], np.column_stack((x0_guess, xf_guess))) dircol.SetInitialTrajectory(PiecewisePolynomial(), initial_x_trajectory) # Solve direct collocation result = Solve(dircol) assert result.is_success() print("Found a solution!") # PLOTTING N_plot = 200 # Plot trajectory x_trajectory = dircol.ReconstructStateTrajectory(result) times = np.linspace(x_trajectory.start_time(), x_trajectory.end_time(), N_plot) x_knots = np.hstack([x_trajectory.value(t) for t in times]) z = x_knots[3, :] x = x_knots[4, :] y = x_knots[5, :] plot_trj_3_wind(np.vstack((x, y, z)).T, dir_vector) # Plot input u_trajectory = dircol.ReconstructInputTrajectory(result) u_knots = np.hstack([u_trajectory.value(t) for t in times]) plot_input_zhao_glider(times, u_knots.T) plt.show() return 0
def make_real_dircol_mp(expmt="cartpole", seed=1776): global tree global plant global context global dircol # TODO: use the seed in some meaningful way: # https://github.com/RobotLocomotion/drake/blob/master/systems/stochastic_systems.h assert expmt in ("cartpole", "acrobot") # expmt = "cartpole" # State: (x, theta, x_dot, theta_dot) Input: x force # expmt = "acrobot" # State: (theta1, theta2, theta1_dot, theta2_dot) Input: Elbow torque if expmt == "cartpole": tree = RigidBodyTree("/opt/underactuated/src/cartpole/cartpole.urdf", FloatingBaseType.kFixed) plant = RigidBodyPlant(tree) else: tree = RigidBodyTree("/opt/underactuated/src/acrobot/acrobot.urdf", FloatingBaseType.kFixed) plant = AcrobotPlant() context = plant.CreateDefaultContext() if expmt == "cartpole": dircol = DirectCollocation(plant, context, num_time_samples=21, minimum_timestep=0.1, maximum_timestep=0.4) else: dircol = DirectCollocation(plant, context, num_time_samples=21, minimum_timestep=0.05, maximum_timestep=0.2) dircol.AddEqualTimeIntervalsConstraints() if expmt == "acrobot": # Add input limits. torque_limit = 8.0 # N*m. u = dircol.input() dircol.AddConstraintToAllKnotPoints(-torque_limit <= u[0]) dircol.AddConstraintToAllKnotPoints(u[0] <= torque_limit) initial_state = (0., 0., 0., 0.) dircol.AddBoundingBoxConstraint(initial_state, initial_state, dircol.initial_state()) # More elegant version is blocked on drake #8315: # dircol.AddLinearConstraint(dircol.initial_state() == initial_state) if expmt == "cartpole": final_state = (0., math.pi, 0., 0.) else: final_state = (math.pi, 0., 0., 0.) dircol.AddBoundingBoxConstraint(final_state, final_state, dircol.final_state()) # dircol.AddLinearConstraint(dircol.final_state() == final_state) # R = 10 # Cost on input "effort". # u = dircol.input() # dircol.AddRunningCost(R*u[0]**2) # Add a final cost equal to the total duration. dircol.AddFinalCost(dircol.time()) initial_x_trajectory = \ PiecewisePolynomial.FirstOrderHold([0., 4.], np.column_stack((initial_state, final_state))) dircol.SetInitialTrajectory(PiecewisePolynomial(), initial_x_trajectory) return dircol, tree
def _solve_traj_opt(self, initial_state, constrain_final_state=True, duration_bounds=None, d=0.0, verbose=False): '''Finds a trajectory from an initial state, optionally to a final state. Args: initial_state (tuple): the initial state final_state (tuple): the final state (default to None, final state unconstrained) duration (tuple): the min and max duration of the trajectory (default to None, no duration constraints) d (float): constant disturbance force verbose (bool): enables/disables verbose output Returns: pydrake.trajectories.PiecewisePolynomial: the planned trajectory float: the cost of the planned trajectory Raises: RuntimeError: raised if the optimization fails ''' print("Initial state: {}\nFinal state: {}\nMin duration: {} s\nMax duration: {} s".format( initial_state, constrain_final_state, duration_bounds[0], duration_bounds[1])) traj_opt = DirectCollocation(self.plant, self.context, self.opt_params['num_time_samples'], self.opt_params['minimum_timestep'], self.opt_params['maximum_timestep']) traj_opt.AddEqualTimeIntervalsConstraints() # Add bounds on the total duration of the trajectory if duration_bounds: traj_opt.AddDurationBounds(duration_bounds[0], duration_bounds[1]) # TODO make input limits a paramter limits_low = [-15., -15.] limits_upp = [15., 15.] x = traj_opt.state() u = traj_opt.input() t = traj_opt.time() # TODO assuming disturbance is at the last index for i in range(len(u) - 1): traj_opt.AddConstraintToAllKnotPoints(limits_low[i] <= u[i]) traj_opt.AddConstraintToAllKnotPoints(u[i] <= limits_upp[i]) traj_opt.AddConstraintToAllKnotPoints(u[len(u) - 1] == d) for signed_dist_func in self.signed_dist_funcs: traj_opt.AddConstraintToAllKnotPoints(signed_dist_func(x) >= 0) traj_opt.AddRunningCost(traj_opt.timestep(0) * self.running_cost(x, u, t)) traj_opt.AddFinalCost(self.final_cost(x, u, t)) # Add initial and final state constraints traj_opt.AddBoundingBoxConstraint(initial_state, initial_state, traj_opt.initial_state()) if self.final_state_constraint and constrain_final_state: traj_opt.AddConstraint(self.final_state_constraint(traj_opt.final_state()) == 0) # # TODO this is redundant with the final state equality constraint above # if final_state: # traj_opt.AddBoundingBoxConstraint(final_state, final_state, # traj_opt.final_state()) # initial_x_trajectory = PiecewisePolynomial.FirstOrderHold([0., 0.4 * 21], # np.column_stack((initial_state, # final_state))) # traj_opt.SetInitialTrajectory(PiecewisePolynomial(), initial_x_trajectory) # else: # initial_x_trajectory = PiecewisePolynomial.FirstOrderHold([0., 0.4 * 21], # np.column_stack((initial_state, # initial_state))) # traj_opt.SetInitialTrajectory(PiecewisePolynomial(), initial_x_trajectory) initial_x_trajectory = PiecewisePolynomial.FirstOrderHold([0., 0.4 * 21], np.column_stack((initial_state, initial_state))) traj_opt.SetInitialTrajectory(PiecewisePolynomial(), initial_x_trajectory) result = traj_opt.Solve() if result != SolutionResult.kSolutionFound: raise RuntimeError('Direct collocation failed from initial state {}!'.format(initial_state)) state_samples = traj_opt.GetStateSamples() input_samples = traj_opt.GetInputSamples() time_samples = traj_opt.GetSampleTimes() # for debugging hs = [time_samples[i+1] - time_samples[i] for i in range(len(time_samples)) if i < len(time_samples) - 1] #print(hs) total_cost = 0. for k in range(state_samples.shape[1]): total_cost += (hs[0] * self.running_cost(state_samples[:, k], input_samples[:, k], time_samples[k])) if verbose: for i, phi in enumerate(self.signed_dist_funcs): print("\tsigned dist {}: {}".format(i, signed_dist_func(state_samples[:, k]))) if verbose: print("Total cost is {}".format(total_cost)) u_traj = traj_opt.ReconstructInputTrajectory() times = np.linspace(u_traj.start_time(), u_traj.end_time(), 100) u_lookup = np.vectorize(lambda t: u_traj.value(t)[0]) u_values = u_lookup(times) plt.figure() plt.plot(times, u_values) plt.xlabel('time (seconds)') plt.ylabel('force (Newtons)') plt.show() return traj_opt.ReconstructStateTrajectory(), total_cost