Error status 3: ACADOS_MINSTEP. Any configuration to alleviate it?

roangel · September 30, 2021, 12:33pm

Hi

I am using python to generate C code for a contouring controller using HPIPM, and then use that C code in a C++ environment. However, sometimes I get this error:

which translates to ACADOS_MINSTEP error, if I am not mistaken.

Now, the question I have is: I am using SQP_RTI, why does it say that there are multiple iterations? I thought only one iteration was used. Also, is there any option that I can modify in the nlp solver to alleviate the occurrence of this error? Or is it inherently linked to the complexity of the cost function I am working with?

Thanks,

Angel.

FreyJo · October 1, 2021, 8:28am

Hi Angel,

In RTI, a single QP is solved within every solver call.
However, the QP solver itself (here HPIPM) is also an iterative method.
Error status 3 in HPIPM corresponds to “NaN in solution detected”.

github.com

giaf/hpipm/blob/8fbddb7e9bee0befbe1915e0651bf23cf99c841b/include/hpipm_common.h#L62


      
          	SPEED, // focus on speed, relative IPM formulation
          	BALANCE, // balanced mode, relative IPM formulation
          	ROBUST, // focus on robustness, relative IPM formulation
          	};
          
          enum hpipm_status
          	{
          	SUCCESS, // found solution satisfying accuracy tolerance
          	MAX_ITER, // maximum iteration number reached
          	MIN_STEP, // minimum step length reached
          	NAN_SOL, // NaN in solution detected
          	INCONS_EQ, // unconsistent equality constraints
          	};
          
          
          
          #ifdef __cplusplus
          } /* extern "C" */
          #endif

I am not really sure why this happens and don’t know your problem formulation.
Maybe regularizing the Hessian can help.
https://docs.acados.org/python_api/#acados_template.acados_ocp.AcadosOcpOptions.levenberg_marquardt

Best,
Jonathan

roangel · October 1, 2021, 12:10pm

Hi again,

Thanks a lot! this solution does in fact solve my problem! The cost function is the one in this paper ([2108.13205] Model Predictive Contouring Control for Near-Time-Optimal Quadrotor Flight), Eq. (16).

For me to understand better, is the levenberg_marquardt term the lambda in this paper https://people.duke.edu/~hpgavin/ce281/lm.pdf Eq. (12)? Is there any way to choose a good value?

Thanks again!

Best,
Angel.

FreyJo · October 14, 2021, 8:04am

Hi,

sorry for the late reply.

It is implemented here:

github.com

acados/acados/blob/acbd276e9622cf53e0b2705169f5fcf710cb7dd4/acados/ocp_nlp/ocp_nlp_common.c#L1959-L1976

    
      
          if (i < N)
          {
              // Levenberg Marquardt term: Ts[i] * levenberg_marquardt * eye()
              if (opts->levenberg_marquardt > 0.0)
                  blasfeo_ddiare(nu[i] + nx[i], in->Ts[i] * opts->levenberg_marquardt,
                                 mem->qp_in->RSQrq+i, 0, 0);
          
          
    // dynamics
              config->dynamics[i]->update_qp_matrices(config->dynamics[i], dims->dynamics[i],
                      in->dynamics[i], opts->dynamics[i], mem->dynamics[i], work->dynamics[i]);
          }
          else
          {
              // Levenberg Marquardt term: 1.0 * levenberg_marquardt * eye()
              if (opts->levenberg_marquardt > 0.0)
                  blasfeo_ddiare(nu[i] + nx[i], opts->levenberg_marquardt,
                                 mem->qp_in->RSQrq+i, 0, 0);
          }

It works like in the paper you linked, but the factor is then multiplied with the time step length for the intermediate stages.

Best,
Jonathan