Do you see a problem here? I get exact new results for both runs. MU = 0.5 # Coefficient…
- User
Do you see a problem here? I get exact new results for both runs.
MU = 0.5 # Coefficient of friction T = 1.5 # Total time T_UB = 2.0 # Upper bound on total time N_knots = 25 # Number of knot points per segment M = 74 # Mass of the pendulum (kg) L = 1.74 # Length of the pendulum (m) lleg = Link(m=0.093 * M, l=0.246 * L) # Lower leg link uleg = Link(m=0.200 * M, l=0.245 * L) # Upper leg link torso = Link(m=0.678 * M, l=0.47 * L) # Torso link
lleg = Link(m=1, l=1) # Link 1
uleg = Link(m=1, l=1) # Link 2
torso = Link(m=1, l=1) # Link 3
SEAT_POS = (-0.4, 0.4) # Seat position FOOT_X_LIMS = (-0.05, 0.2) # Foot x limits R_weights_init = (1, 1, 1) # Control weights R_weights = (1, 3, 1)
U_UB = (None,) * 3 # Upper bound on control inputs U_LB = (None,) * 3 # Lower bound on control inputs
Define the system
sts = STS(link1=lleg, link2=uleg, link3=torso, mu=MU, seat_pos=SEAT_POS, foot_x_lims=FOOT_X_LIMS) sts_optimizer = STSTrajectoryOptimizer(sts, control_weights=R_weights_init, )
x0 = (None, None, -0.01, 0, 0, 0) # Initial state xf = (0, 0, 0, 0, 0, 0)
j1, j2 = solve_ik_optimization(*SEAT_POS, lleg.l, uleg.l)
Adjust the joint angles to match STS convention
j2 = j1 + j2 - np.pi/2 j1 = j1 - np.pi/2 print(j1, j2) sit_mode_initial_guess_x0 = np.array([j1, j2, -0.01, 0, 0, 0]).reshape(-1, 1) sit_mode_initial_guess_xf = np.array([j1, j2, -np.pi/6, 0, 0, -0.1]).reshape(-1, 1) sit_mode_initial_guess = np.hstack((sit_mode_initial_guess_x0, sit_mode_initial_guess_xf))
mode_sit = Mode( name='sit', dynamics_fn=sts.sit_dynamics_fn, n_x=6, n_u=3, constraints_fns=[ sts_optimizer.seat_contact_no_slip_constraint_fn, sts_optimizer.sit_path_constraint_fn, ], initial_x_guess=sit_mode_initial_guess, u_lb=U_LB, u_ub=U_UB, )
mode_stand = Mode( name='stand', dynamics_fn=sts.stand_dynamics_fn, n_x=6, n_u=3, constraints_fns=[
# sts_optimizer.COP_constraint_fn, sts_optimizer.knee_joint_constraint_fn, ], u_lb=U_LB, u_ub=U_UB, )transition_sit_stand = Transition( guard_fn=lambda x, u: sts.sit_contact_force_fn(x, u)[1], # y component of the contact force reset_map_fn=sts_optimizer.sit_stand_map_fn, )
hybrid_params = HybridSystemParams( mode_sequence=[mode_sit, mode_stand], transitions=[transition_sit_stand], x0=x0, # Initial state xf=xf, # Final state # T = T, # Total time T_ub = T_UB, # Upper bound on total time min_time_step = 1E-6, num_knot_points_per_segment = N_knots, # Number of knot points per segment )
hybrid_optimizer = HybridTrajectoryOptimizer( params=hybrid_params, cost_fn=sts_optimizer.cost_function, global_constraints_fns=[sts_optimizer.base_contact_no_slip_global_constraint_fn, sts_optimizer.COP_global_constraint_fn, sts_optimizer.equal_time_segment_global_constraint_fn, sts_optimizer.control_effort_vanishes_global_constraint_fn, ], )
Solve the optimization problem
solver_name = 'ipopt' solver_options = { 'ipopt.print_level': 5, 'ipopt.max_iter': 3000, 'ipopt.fixed_variable_treatment': 'relax_bounds', 'ipopt.tol': 1E-3, }
solution = hybrid_optimizer.solve(solver_name=solver_name, opts=solver_options)
Now use the solution to warm start the simulation
DIRTY JOB
sts_optimizer.R = ca.diag(R_weights) # Update the control weights
hybrid_optimizer._setup_objective() # Re-setup the objective function with the new weights
solution = hybrid_optimizer.solve(solver_name, solver_options, warm_start=True)
SEAT_POS_NEW = (None, 0.2) # New seat position
Update the initial guess for the sit mode
mode_sit.initial_x_guess = solution['states'][0] mode_stand.initial_x_guess = solution['states'][1] sts_new = STS(link1=lleg, link2=uleg, link3=torso, mu=MU, seat_pos=SEAT_POS_NEW, foot_x_lims=FOOT_X_LIMS) sts_optimizer_new = STSTrajectoryOptimizer(sts_new, control_weights=R_weights_init, ) hybrid_params_new = HybridSystemParams( mode_sequence=[mode_sit, mode_stand], transitions=[transition_sit_stand], x0=x0, # Initial state xf=xf, # Final state # T = T, # Total time T_ub=T_UB, # Upper bound on total time min_time_step=1E-6, num_knot_points_per_segment=N_knots, # Number of knot points per segment )
Recreate the hybrid optimizer with the new system
hybrid_optimizer_new = HybridTrajectoryOptimizer( params=hybrid_params_new, cost_fn=sts_optimizer_new.cost_function, global_constraints_fns=[sts_optimizer_new.base_contact_no_slip_global_constraint_fn, sts_optimizer_new.COP_global_constraint_fn, sts_optimizer_new.equal_time_segment_global_constraint_fn, sts_optimizer_new.control_effort_vanishes_global_constraint_fn, ], )
Solve the optimization problem again with the new system
solution = hybrid_optimizer_new.solve(solver_name=solver_name, opts=solver_options)
End of dirty job
- Assistant
Let's break down the code and identify potential issues that could lead to getting exact new results for both runs, which seems unexpected given that you are changing the
SEAT_POSbetween the two runs.Potential Issues and Suggestions
-
Seat Position Update:
- Issue:
SEAT_POS_NEW = (None, 0.2)might not be correctly updating the seat position. The first element isNone, which could be causing an issue. - Suggestion: Ensure that
SEAT_POS_NEWis a valid tuple with two numeric values. For exampleSEAT_POS_NEW = (-0.4, 0.2)
SEAT_POS_NEW = (-0.4, 0.2) # New seat position - Issue:
-
Initial Guess Update:
- Issue: You are updating the initial guess for
mode_sitandmode_standusing the solution from the previous optimisation. However you are not updating the initial guess for the new seat position which might lead to the same result. - Suggestion: Update the initial joint angles
j1andj2for the new seat position.
# Update joint angles for the new seat position j1_new, j2_new = solve_ik_optimization(*SEAT_POS_NEW, lleg.l, uleg.l) # Adjust the joint angles to match STS convention j2_new = j1_new + j2_new - np.pi/2 j1_new = j1_new - np.pi/2 print(j1_new, j2_new) sit_mode_initial_guess_x0_new = np.array([j1_new, j2_new, -0.01, 0, 0, 0]).reshape(-1, 1) sit_mode_initial_guess_xf_new = np.array([j1_new, j2_new, -np.pi/6, 0, 0, -0.1]).reshape(-1, 1) sit_mode_initial_guess_new = np.hstack((sit_mode_initial_guess_x0_new, sit_mode_initial_guess_xf_new)) # Update the initial guess for the sit mode with the new initial guess mode_sit.initial_x_guess = sit_mode_initial_guess_new - Issue: You are updating the initial guess for
-
Control Weights:
- Issue: You initially set
R_weights_init = (1, 1, 1)and then comment out the code that updates the control weights toR_weights = (1, 3, 1). This means the control weights are never actually updated. - Suggestion: Uncomment and correctly update the control weights before resolving the optimisation problem.
# Update the control weights sts_optimizer_new.R = np.diag(R_weights) # Update the control weights hybrid_optimizer_new._setup_objective() # Re-setup the objective function with the new weights - Issue: You initially set
-
Constraint Functions:
- Issue: Ensure that the constraint functions are correctly defined and are sensitive to changes in
SEAT_POS. If the constraints do not depend onSEAT_POS, changing it will not affect the results - Suggestion: Verify that
seat_contact_no_slip_constraint_fn,sit_path_constraint_fn, and other constraint functions correctly incorporateSEAT_POS.
- Issue: Ensure that the constraint functions are correctly defined and are sensitive to changes in
-
Optimizer Settings:
- Issue: The solver settings might be too aggressive or not sensitive enough to detect changes.
- Suggestion: Adjust the solver tolerance and print level to ensure the solver is not terminating prematurely.
solver_options = { 'ipopt.print_level': 5, 'ipopt.max_iter': 5000, # Increase the maximum iterations 'ipopt.fixed_variable_treatment': 'relax_bounds', 'ipopt.tol': 1E-6, # Decrease the tolerance for more precision }
Updated Code Section
Here is the updated section of the code incorporating the suggestions:
# Update the seat position SEAT_POS_NEW = (-0.4, 0.2) # New seat position # Update joint angles for the new seat position j1_new, j2_new = solve_ik_optimization(*SEAT_POS_NEW, lleg.l, uleg.l) # Adjust the joint angles to match STS convention j2_new = j1_new + j2_new - np.pi/2 j1_new = j1_new - np.pi/2 print(j1_new, j2_new) sit_mode_initial_guess_x0_new = np.array([j1_new, j2_new, -0.01, 0, 0, 0]).reshape(-1, 1) sit_mode_initial_guess_xf_new = np.array([j1_new, j2_new, -np.pi/6, 0, 0, -0.1]).reshape(-1, 1) sit_mode_initial_guess_new = np.hstack((sit_mode_initial_guess_x0_new, sit_mode_initial_guess_xf_new)) # Update the initial guess for the sit mode with the new initial guess mode_sit.initial_x_guess = sit_mode_initial_guess_new mode_stand.initial_x_guess = solution['states'][1] # Create the new STS system with the updated seat position sts_new = STS(link1=lleg, link2=uleg, link3=torso, mu=MU, seat_pos=SEAT_POS_NEW, foot_x_lims=FOOT_X_LIMS) sts_optimizer_new = STSTrajectoryOptimizer(sts_new, control_weights=R_weights_init, ) # Update the control weights sts_optimizer_new.R = np.diag(R_weights) # Update the control weights hybrid_optimizer_new._setup_objective() # Re-setup the objective function with the new weights # Update the hybrid system parameters with the new system hybrid_params_new = HybridSystemParams( mode_sequence=[mode_sit, mode_stand], transitions=[transition_sit_stand], x0=x0, # Initial state xf=xf, # Final state T_ub=T_UB, # Upper bound on total time min_time_step=1E-6, num_knot_points_per_segment=N_knots, # Number of knot points per segment ) # Recreate the hybrid optimizer with the new system hybrid_optimizer_new = HybridTrajectoryOptimizer( params=hybrid_params_new, cost_fn=sts_optimizer_new.cost_function, global_constraints_fns=[sts_optimizer_new.base_contact_no_slip_global_constraint_fn, sts_optimizer_new.COP_global_constraint_fn, sts_optimizer_new.equal_time_segment_global_constraint_fn, sts_optimizer_new.control_effort_vanishes_global_constraint_fn, ], ) # Adjust solver options for more precision solver_options = { 'ipopt.print_level': 5, 'ipopt.max_iter': 5000, 'ipopt.fixed_variable_treatment': 'relax_bounds', 'ipopt.tol': 1E-6, } # Solve the optimization problem again with the new system solution = hybrid_optimizer_new.solve(solver_name=solver_name, opts=solver_options)Conclusion
By ensuring the seat position is correctly updated, adjusting the initial guesses for the new configuration, updating the control weights, and tweaking the solver settings, you should see different results between the two runs as expected.
-
Other models, same or similar prompt
Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.