Hi, Fenglong
It’s indeed that if you use pinocchio3-casadi interface, the code generation and compilation with C++ takes much more time. I’ve tried to do so(centroidal dynamics things) the same as what ocs2 did and it could keep the robot standing if I didn’t add too complex constaint. But the solver failed when I add non-velocity constraint to the stance leg with pinocchio. I shared the problem here https://discourse.acados.org/t/trouble-with-nonliear-constraints-on-derivative-of-state-sqp-rti-qp-solver-returned-error-status-3-qp-iteration-15/1487?u=lrc, but there’s no solution yet. I preliminary think that could be solved by provide external hessain and gradient of the dynamics and constraint using pinocchio for acados, but it’s too hard for me to do so. So I used ocs2 instead and it was really efficient in my another simulation. What I did with acados I mentioned before was a simplified version of that using single rigid body dynamics of the COM and substituting the joint states with foot position and it had similar performance. 
What I’m interested in now was to generate a heuristic motion for quadruped robot without a given contact sequence, as it’s more intuitive for an animal’s motion. Older work on it was considering the feasible impulse set(FIS) which assess the contribution of each leg comparing their feasible impulse polytope(FIP) across the horizon(N steps), and it generate the foothold and contact sequences. While the benefit of this is that there’s no need to modify the well-tested mpc-wbc framework, there’s no room for further research. So I’m researching for ocp based methods considering contact-implicit model to generate the gait automatically. Most methods are hard to implement because of the calculation cost, while some more efficient methods are too hard to try it myself. It’s the first year for me as a graduate student and I know too little about this, but I’m going to keep on it. Thanks for sharing and looking forward to more future exchanges with you! 
Best,
lrc