Skip to content

Backends

A backend adapts one trajectory-optimization library to blendmpc.core.MPCPolicy. The contract:

  • implement solve(x0, us_init, xs_init) -> MPCSolution — one open-loop OCP from x0, warm-started when trajectories are given;
  • the existing blend test suite must pass over the new backend unchanged.

Crocoddyl — blendmpc.solvers.crocoddyl.CrocoddylMPC

CrocoddylMPC(problem_factory,           # x0 -> crocoddyl.ShootingProblem
             solver_cls=SolverBoxFDDP,  # honors control limits
             max_iter=5,                # warm-started iterations per step
             max_iter_first=100,        # cold-solve budget
             u_init=None,               # symmetry-breaking cold init
             multistart_iter=0)         # optional per-step cold restart

Installs from PyPI (pip install crocoddyl). FDDP's feasibility-driven structure is strong on swing-up-like problems; the box solvers enforce control bounds inside the backward pass.

acados — blendmpc.solvers.acados.AcadosMPC

AcadosMPC(ocp_factory,        # x0 -> AcadosOcp (CasADi model, EXTERNAL cost)
          build_dir=None,     # where generated C code is compiled
          u_init=None)        # symmetry-breaking cold init

Needs the acados C library built from source (ACADOS_SOURCE_DIR, LD_LIBRARY_PATH). The first solve generates and compiles problem-specific C code into build_dir; reset() keeps the compiled solver and only clears the warm start. See blendmpc.envs.pendulum_acados for a complete AcadosOcp factory using discrete dynamics and an external cost.

Same OCP, different local solutions

On the pendulum swing-up, FDDP (Crocoddyl) and SQP (acados) solve the same OCP — verified by cross-seeding their solutions — but from some initial states SQP settles on a one-extra-swing route where FDDP goes direct. Local solvers differ in which basin they find, not just in speed. If a backend "underperforms", seed it with the other backend's solution before concluding the adapter is wrong.

Writing a new backend

  1. Subclass MPCPolicy, implement solve(). Map your solver's warm-start API to us_init/xs_init and return an MPCSolution (put solver diagnostics in info).
  2. Decide what reset() costs: rebuild cheap per-episode state, keep expensive compiled artifacts.
  3. Copy tests/test_acados_backend.py as a template: interface parity, torque limits respected, stabilization, zero-residual-equals-MPC through ResidualMPCEnv.