Robotics algorithms typically depend on various parameters, the choice ofwhich significantly affects the robot's performance. While an initial guess forthe parameters may be obtained from dynamic models of the robot, parameters areusually tuned manually on the real system to achieve the best performance.Optimization algorithms, such as Bayesian optimization, have been used toautomate this process. However, these methods may evaluate parameters duringthe optimization process that lead to safety-critical system failures.Recently, a safe Bayesian optimization algorithm, called SafeOpt, has beendeveloped and applied in robotics, which guarantees that the performance of thesystem never falls below a critical value; that is, safety is defined based onthe performance function. However, coupling performance and safety is notdesirable in most cases. In this paper, we define separate functions forperformance and safety. We present a generalized SafeOpt algorithm that, givenan initial safe guess for the parameters, maximizes performance but onlyevaluates parameters that satisfy all safety constraints with high probability.It achieves this by modeling the underlying and unknown performance andconstraint functions as Gaussian processes. We provide a theoretical analysisand demonstrate in experiments on a quadrotor vehicle that the proposedalgorithm enables fast, automatic, and safe optimization of tuning parameters.Moreover, we show an extension to context- or environment-dependent, safeoptimization in the experiments.
展开▼
