Rollout system trajectories with benchmarked solvers themselves

[stephane-caron]

Current context

MPC problems are generated statically using a reference solver (e.g. quadprog for LIPMWALK problems, ProxQP for WHLIPBAL problems), saved to data/ and then run with benchmark solvers.

Describe the solution you'd like

This comparison is not fair in the context of MPC: in MPC, when a solver fails to find a solution, the integrator moves on with the previous one, often leading to "failure in series" (loss of recursive feasibility). The benchmark would be more realistic (more relevant to roboticists) if we used solver solutions to rollout trajectories.

See also

• Cold start only qpbenchmark#101