Segfault setting a problem with no nodes

[Stéphane Caron]

The API says we can setup a solver for a problem. This makes me immediately want to try setting up different problems in succession 😜

    solver.setup(problem)  # sneaky line
    for i in range(nb_nodes):
        stage = proxddp.StageModel(running_cost, integrator)
        problem.addStage(stage)
    solver.setup(problem)

On my machine this leads to:

$ python example.py  # with sneaky line
[1]    87811 segmentation fault (core dumped)  python trajopt.py

$ python example.py  # without sneaky line
iter｜  alpha   ｜inner_crit｜ prim_err ｜ dual_err ｜   xreg   ｜  dphi0   ｜  merit   ｜ delta_M  ｜al_iter
Successfully converged.

→ Suggested unit test.

Reproduction code

Just a part of the cartpole.py example. Sorry this is not minimal:

import numpy as np
import pinocchio as pin
import proxddp
import proxnlp
from robot_descriptions.loaders.pinocchio import load_robot_description

if __name__ == "__main__":
    full_robot = load_robot_description(
        "upkie_description",
        root_joint=pin.JointModelFreeFlyer(),
    )
    robot = full_robot.buildReducedRobot(
        list_of_joints_to_lock=[
            f"{side}_{joint}"
            for side in ("left", "right")
            for joint in ("hip", "knee")
        ]
    )

    model = robot.model
    timestep = 0.01  # [s]
    control_dim = 1
    actuation_matrix = np.array([0.0] * 6 + [+1.0, -1.0]).reshape(
        (model.nv, control_dim)
    )

    phase_space = proxnlp.manifolds.MultibodyPhaseSpace(model)
    continuous_dynamics = proxddp.dynamics.MultibodyFreeFwdDynamics(
        phase_space, actuation_matrix
    )
    integrator = proxddp.dynamics.IntegratorSemiImplEuler(
        continuous_dynamics, timestep
    )

    # Control running cost
    wu = 1e-4 * np.ones(control_dim)
    control_reg_cost = proxddp.QuadraticResidualCost(
        proxddp.ControlErrorResidual(phase_space.ndx, np.zeros(control_dim)),
        timestep * np.diag(wu),
    )

    # Target running cost
    target = pin.SE3.Identity()
    target.translation = np.array([0.5, 0.0, 0.0])
    target_weights = np.zeros(6)
    target_weights[:3] = np.ones(3) * 1.0
    frame_residual = proxddp.FramePlacementResidual(
        phase_space.ndx,
        control_dim,
        model,
        target,
        model.getFrameId("chassis"),
    )
    target_cost = proxddp.QuadraticResidualCost(
        frame_residual,
        np.diag(target_weights) * timestep,
    )

    # Running cost stack
    running_cost = proxddp.CostStack(phase_space.ndx, control_dim)
    running_cost.addCost(control_reg_cost)
    running_cost.addCost(target_cost)

    # Terminal cost
    # XXX: warning, re-using target cost here! If it breaks start there
    terminal_cost = proxddp.CostStack(phase_space.ndx, control_dim)
    terminal_cost.addCost(target_cost)

    nb_nodes = 100
    initial_state = phase_space.neutral()
    problem = proxddp.TrajOptProblem(
        initial_state, control_dim, phase_space, terminal_cost
    )

    mu_init = 4e-2
    rho_init = 1e-2
    verbose = proxddp.VerboseLevel.VERBOSE
    TOL = 1e-4
    MAX_ITER = 200
    solver = proxddp.SolverProxDDP(
        TOL, mu_init, rho_init=rho_init, max_iters=MAX_ITER, verbose=verbose
    )

    solver.setup(problem)
    for i in range(nb_nodes):
        stage = proxddp.StageModel(running_cost, integrator)
        problem.addStage(stage)
    solver.setup(problem)

    u0 = np.zeros(control_dim)
    controls = [u0] * nb_nodes
    states = proxddp.rollout(integrator, initial_state, controls)
    solver.run(problem, states, controls)

[Wilson]

Nice catch ! What do you get in debug mode ? Is it the first setup() which produces the segfault? That would be my guess since the problem is "empty" (e.g. has no other nodes)

[Wilson]

Also, you're welcome to add the example to the repo in a PR :)

[Stéphane Caron]

Is it the first setup() which produces the segfault?

Whoops, my bad issue description. You're 🎯, the first setup() is the segfault-guilty one.

Also, you're welcome to add the example to the repo in a PR :)

Way to go 👍 #47