Comment by benchmarkist

Neural networks use smooth manifolds as their underlying inductive bias so in theory it should be possible to incorporate smooth kinematic and Hamiltonian constraints but I am certain no one at OpenAI actually understands enough of the theory to figure out how to do that.