GTSAM suffers from having created many different "custom methods" in FactorGraph to add factors. It would be much nicer to be as close as possible to a std::vector. Since the hybrid API is still in complete beta, I am proposing to radically simplify, starting with HybridBayesNet, by getting rid of the non-standard methods in favor of:

• hbn.emplace_back(new GaussianMixture...));
• hbn.emplace_back(new GaussianConditional...));
• hbn.emplace_back(new DiscreteConditional...));

Adding shared pointers also works, should you need them somewhere else:

• hbn.push_back(shared_ptr_to_a_conditional);

Note, in python things are automatically shared pointers, so there we always use push_back.

Similarly, we just use at:

• hbn.at(i).asMixture();
• hbn.at(i).asGaussian();
• hbn.at(i).asDiscrete();

@ProfFan was very helpful making this work.