I don't claim any insider knowledge about how specifically BD is doing it. But using sampling-based verification is quite routine in the verification community... especially with neural lyapunov/barrier functions these days. For a view from Lyapunov / worst-case analysis, we explored one particular version back in the day: groups.csail.mit.edu/robotics-center/public_papers/Reist10.pdf
at 59:41 you say that Boston Dynamics wouldn't use SOS, but instead use empirical. Do you discuss building a funnel from empirical evidence anywhere?
I don't claim any insider knowledge about how specifically BD is doing it. But using sampling-based verification is quite routine in the verification community... especially with neural lyapunov/barrier functions these days. For a view from Lyapunov / worst-case analysis, we explored one particular version back in the day: groups.csail.mit.edu/robotics-center/public_papers/Reist10.pdf